detection and segmentation
play

Detection and Segmentation CS60010: Deep Learning Abir Das IIT - PowerPoint PPT Presentation

Aug 05, 2023 •249 likes •1.41k views

Detection and Segmentation CS60010: Deep Learning Abir Das IIT Kharagpur March 04 and 05, 2020 Detection RCNN Architectures YOLO Segmentation Agenda To get introduced to two important tasks of computer vision - detection and segmentation

  1. Detection RCNN Architectures YOLO Segmentation Fast R-CNN Fast R-CNN source. Reproduced with permission. Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 2018 63 § R Girshick, ‘Fast R-CNN’, ICCV 2015 CS231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 27 / 106

  2. Detection RCNN Architectures YOLO Segmentation Fast R-CNN Fast R-CNN source. Reproduced with permission. Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 2018 63 CS231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 28 / 106

  3. Detection RCNN Architectures YOLO Segmentation Fast R-CNN Fast R-CNN source. Reproduced with permission. Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 2018 63 CS231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 29 / 106

  4. Detection RCNN Architectures YOLO Segmentation Fast R-CNN Fast R-CNN: RoI Pooling Divide projected Project proposal proposal into 7x7 onto features Fully-connected grid, max-pool within each cell layers CNN Hi-res input image: Hi-res conv features: RoI conv features: Fully-connected layers expect 3 x 640 x 480 512 x 20 x 15; 512 x 7 x 7 low-res conv features: with region for region proposal 512 x 7 x 7 proposal Projected region proposal is e.g. 512 x 18 x 8 (varies per proposal) CS231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 30 / 106

  5. Detection RCNN Architectures YOLO Segmentation Fast R-CNN Fast R-CNN CS231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 31 / 106

  6. Detection RCNN Architectures YOLO Segmentation Fast R-CNN Fast R-CNN (Training) CS231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 32 / 106

  7. Detection RCNN Architectures YOLO Segmentation Fast R-CNN R-CNN vs SPP vs Fast R-CNN Girshick et al, “Rich feature hierarchies for accurate object detection and semantic segmentation”, CVPR 2014. He et al, “Spatial pyramid pooling in deep convolutional networks for visual recognition”, ECCV 2014 Girshick, “Fast R-CNN”, ICCV 2015 Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 2018 70 CS231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 33 / 106

  8. Detection RCNN Architectures YOLO Segmentation Fast R-CNN R-CNN vs SPP vs Fast R-CNN Problem : Runtime dominated by region proposals! Girshick et al, “Rich feature hierarchies for accurate object detection and semantic segmentation”, CVPR 2014. He et al, “Spatial pyramid pooling in deep convolutional networks for visual recognition”, ECCV 2014 Girshick, “Fast R-CNN”, ICCV 2015 Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 2018 70 CS231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 34 / 106

  9. Detection RCNN Architectures YOLO Segmentation Fast R-CNN Region proposals Feature extraction Classifier Region Proposals: Selective Search Feature Extraction: CNN Pre 2012 Classifier: CNN RCNN Fast RCNN CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 35 / 106

  10. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § The bulk of the time at test time of Fast RCNN is dominated by the region proposal generation. § As Fast RCNN saved computation by sharing the feature generation for all proposals, can some sort of sharing of computation be done for generating region proposals? § The solution is to use the same CNN for region proposal generation too. Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 36 / 106

  11. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § The bulk of the time at test time of Fast RCNN is dominated by the region proposal generation. § As Fast RCNN saved computation by sharing the feature generation for all proposals, can some sort of sharing of computation be done for generating region proposals? § The solution is to use the same CNN for region proposal generation too. § S Ren and K He and R Girshick and J Sun, ‘Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks’, NIPS 2015 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 36 / 106

  12. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § The bulk of the time at test time of Fast RCNN is dominated by the region proposal generation. § As Fast RCNN saved computation by sharing the feature generation for all proposals, can some sort of sharing of computation be done for generating region proposals? § The solution is to use the same CNN for region proposal generation too. § S Ren and K He and R Girshick and J Sun, ‘Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks’, NIPS 2015 § The region proposal generation part is termed as the Region Proposal Network (RPN) Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 36 / 106

  13. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § The bulk of the time at test time of Fast RCNN is dominated by the region proposal generation. § As Fast RCNN saved computation by sharing the feature generation for all proposals, can some sort of sharing of computation be done for generating region proposals? § The solution is to use the same CNN for region proposal generation too. § S Ren and K He and R Girshick and J Sun, ‘Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks’, NIPS 2015 § The region proposal generation part is termed as the Region Proposal Network (RPN) Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 36 / 106

  14. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § The bulk of the time at test time of Fast RCNN is dominated by the region proposal generation. § As Fast RCNN saved computation by sharing the feature generation for all proposals, can some sort of sharing of computation be done for generating region proposals? § The solution is to use the same CNN for region proposal generation too. § S Ren and K He and R Girshick and J Sun, ‘Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks’, NIPS 2015 § The region proposal generation part is termed as the Region Proposal Network (RPN) Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 36 / 106

  15. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § The RPN works as follows: ◮ A small 3x3 conv layer is applied on the last layer of the base conv-net ◮ it produces activation feature map of the same size as the base conv-net last layer feature map (7x7x512 in case of VGG base) ◮ At each of the feature positions (7x7=49 for VGG base), a set of bounding boxes (with different scale and aspect ratio) are evaluated for the following two questions given the 512d feature at that position, what is the probability that each of the bounding boxes centered at the position contains an object? (Classification) Given the same 512d feature can you predict the correct bounding box? (Regression) ◮ These boxes are called ‘ anchor boxes ’ Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 37 / 106

  16. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § The RPN works as follows: ◮ A small 3x3 conv layer is applied on the last layer of the base conv-net ◮ it produces activation feature map of the same size as the base conv-net last layer feature map (7x7x512 in case of VGG base) ◮ At each of the feature positions (7x7=49 for VGG base), a set of bounding boxes (with different scale and aspect ratio) are evaluated for the following two questions given the 512d feature at that position, what is the probability that each of the bounding boxes centered at the position contains an object? (Classification) Given the same 512d feature can you predict the correct bounding box? (Regression) ◮ These boxes are called ‘ anchor boxes ’ Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 37 / 106

  17. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § The RPN works as follows: ◮ A small 3x3 conv layer is applied on the last layer of the base conv-net ◮ it produces activation feature map of the same size as the base conv-net last layer feature map (7x7x512 in case of VGG base) ◮ At each of the feature positions (7x7=49 for VGG base), a set of bounding boxes (with different scale and aspect ratio) are evaluated for the following two questions given the 512d feature at that position, what is the probability that each of the bounding boxes centered at the position contains an object? (Classification) Given the same 512d feature can you predict the correct bounding box? (Regression) ◮ These boxes are called ‘ anchor boxes ’ Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 37 / 106

  18. Detection RCNN Architectures YOLO Segmentation Faster R-CNN Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 38 / 106

  19. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § But how do we get the ground truth data to train the RPN. Consider a ground truth object and its corresponding bounding box Input Input Input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 39 / 106

  20. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § But how do we get the ground truth data to train the RPN. Consider a ground truth object and its corresponding bounding box Consider the projection of this image onto the conv5 layer Max-pool Conv Input Input Input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 40 / 106

  21. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § But how do we get the ground truth data to train the RPN. Consider a ground truth object and its corresponding bounding box Consider the projection of this image onto the conv5 layer Consider one such cell in the output Max-pool Conv Input Input Input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 41 / 106

  22. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § But how do we get the ground truth data to train the RPN. Consider a ground truth object and its corresponding bounding box Consider the projection of this image onto the conv5 layer Consider one such cell in the output Max-pool This cell corresponds to a patch in the Conv original image Input Input Input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 42 / 106

  23. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § But how do we get the ground truth data to train the RPN. Consider a ground truth object and its corresponding bounding box Consider the projection of this image onto the conv5 layer Consider one such cell in the output Max-pool This cell corresponds to a patch in the Conv original image Consider the center of this patch Input Input Input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 43 / 106

  24. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § But how do we get the ground truth data to train the RPN. Consider a ground truth object and its corresponding bounding box Consider the projection of this image onto the conv5 layer Consider one such cell in the output Max-pool This cell corresponds to a patch in the Conv original image Consider the center of this patch We consider anchor boxes of different Input Input Input sizes CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 44 / 106

  25. classification Detection RCNN Architectures YOLO Segmentation Faster R-CNN § But how do we get the ground truth data to train the RPN. For each of these anchor boxes, we would want the classifier to predict 1 if this anchor box has a reason-able overlap (IoU > 0.7) with the true grounding box Max-pool Conv Input Input Input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 45 / 106

  26. regression classification Detection RCNN Architectures YOLO Segmentation Faster R-CNN § But how do we get the ground truth data to train the RPN. For each of these anchor boxes, we would want the classifier to predict 1 if this anchor box has a reason- able overlap (IoU > 0.7) with the true grounding box Max-pool Conv Similarly we would want the regres- sion model to predict the true box (red) from the anchor box (pink) Input Input Input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 46 / 106

  27. Detection RCNN Architectures YOLO Segmentation Faster R-CNN Jointly train with 4 losses: 1. RPN classify object / not object 2. RPN regress box coordinates 3. Final classification score (object classes) 4. Final box coordinates Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - CS231n course, Stanford University May 10, 2018 71 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 47 / 106

  28. Detection RCNN Architectures YOLO Segmentation Faster R-CNN § Faster R-CNN based architectures won a lot of challenges including: ◮ Imagenet Detection ◮ Imagenet Localization ◮ COCO Detection ◮ COCO Segmentation Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 48 / 106

  29. Detection RCNN Architectures YOLO Segmentation Faster R-CNN Region proposals Feature extraction Classifier Region Proposals: CNN Feature Extraction: CNN Classifier: CNN Pre 2012 RCNN Fast RCNN Faster RCNN CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 49 / 106

  30. Detection RCNN Architectures YOLO Segmentation YOLO § The R-CNN pipelines separate proposal generation and proposal classification into two separate stages. § Can we have an end-to-end architecture which does both proposal generation and clasification simultaneously? § The solution gives the YOLO (You Only Look Once) architectures. ◮ J Redmon, S Divvala, R Girshick and A Farhadi, ‘You Only Look Once: Unified, Real-Time Object Detection’, CVPR 2016 - YOLO v1 ◮ J Redmon and A Farhadi, ‘YOLO9000: Better, Faster, Stronger’, CVPR 2017 - YOLO v2 ◮ J Redmon and A Farhadi, ‘YOLOv3: An Incremental Improvement’ arXiv preprint 2018 - YOLO v3 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 50 / 106

  31. Detection RCNN Architectures YOLO Segmentation YOLO § The R-CNN pipelines separate proposal generation and proposal classification into two separate stages. § Can we have an end-to-end architecture which does both proposal generation and clasification simultaneously? § The solution gives the YOLO (You Only Look Once) architectures. ◮ J Redmon, S Divvala, R Girshick and A Farhadi, ‘You Only Look Once: Unified, Real-Time Object Detection’, CVPR 2016 - YOLO v1 ◮ J Redmon and A Farhadi, ‘YOLO9000: Better, Faster, Stronger’, CVPR 2017 - YOLO v2 ◮ J Redmon and A Farhadi, ‘YOLOv3: An Incremental Improvement’ arXiv preprint 2018 - YOLO v3 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 50 / 106

  32. Detection RCNN Architectures YOLO Segmentation YOLO § The R-CNN pipelines separate proposal generation and proposal classification into two separate stages. § Can we have an end-to-end architecture which does both proposal generation and clasification simultaneously? § The solution gives the YOLO (You Only Look Once) architectures. ◮ J Redmon, S Divvala, R Girshick and A Farhadi, ‘You Only Look Once: Unified, Real-Time Object Detection’, CVPR 2016 - YOLO v1 ◮ J Redmon and A Farhadi, ‘YOLO9000: Better, Faster, Stronger’, CVPR 2017 - YOLO v2 ◮ J Redmon and A Farhadi, ‘YOLOv3: An Incremental Improvement’ arXiv preprint 2018 - YOLO v3 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 50 / 106

  33. • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell y · · · · c w x h P ( dog ) P ( dog ) S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 51 / 106

  34. • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell y · · · · c w x h For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 52 / 106

  35. • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell y · · · · c w x h For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 53 / 106

  36. • • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell y · · · · c w x h For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object Width of the bounding box containing the true object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 54 / 106

  37. • • • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell y · · · · c w x h For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object Width of the bounding box containing the true object Height of the bounding box containing the true object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 55 / 106

  38. • • • • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell y · · · · c w h x For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object Width of the bounding box containing the true object Height of the bounding box containing the true object Center (x,y) of the bounding box S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 56 / 106

  39. • • • • • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell · · · · c w h x y For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object Width of the bounding box containing the true object Height of the bounding box containing the true object Center (x,y) of the bounding box Probability of the object in the bounding box belonging to the K th class (C values) S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 57 / 106

  40. • • • • • • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell · · · · c w h x y For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object Width of the bounding box containing the true object Height of the bounding box containing the true object Center (x,y) of the bounding box Probability of the object in the bounding box belonging to the Kth class (C values) S × S grid on input The output layer should contain SxSxBx(5+C) elements Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 58 / 106

  41. • • • • • • • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell · · · · c w h x y For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object Width of the bounding box containing the true object Height of the bounding box containing the true object Center (x,y) of the bounding box Probability of the object in the bounding box belonging to the Kth class (C values) S × S grid on input The output layer should contain SxSxBx(5+C) elements However, each grid cell in YOLO predicts only one object even if there are B anchor boxes per cell Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 59 / 106

  42. • • • • • • • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell · · · · c w h x y For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object Width of the bounding box containing the true object Height of the bounding box containing the true object Center (x,y) of the bounding box Probability of the object in the bounding box belonging to the Kth class (C values) S × S grid on input The output layer should contain SxSxBx(5+C) elements The idea is each grid cell tries to make two boundary box predictions to locate a single object Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 60 / 106

  43. • • • • • • • • • Detection RCNN Architectures YOLO Segmentation YOLO Divide an image into S × S grids (S=7) and P ( cow ) P ( cow ) P ( truck ) P ( truck ) consider B (=2) anchor boxes per grid cell · · · · c w h x y For each such anchor box in each cell we are P ( dog ) P ( dog ) interested in predicting 5 + C quantities Probability (confidence) that this anchor box contains a true object Width of the bounding box containing the true object Height of the bounding box containing the true object Center (x,y) of the bounding box Probability of the object in the bounding box belonging to the Kth class (C values) S × S grid on input The output layer should contain SxSxBx(5+C) elements Thus the output layer contains SxSx(Bx5+C) elements Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 61 / 106

  44. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 62 / 106

  45. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 63 / 106

  46. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 64 / 106

  47. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 65 / 106

  48. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 66 / 106

  49. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 67 / 106

  50. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 68 / 106

  51. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 69 / 106

  52. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 70 / 106

  53. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object in it and the type of the object S × S grid on input CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 71 / 106

  54. Detection RCNN Architectures YOLO Segmentation YOLO § During inference/test phase, how do we interpret these S × S × ( B × 5 + C ) outputs? § For each cell we compute the bounding box, its confidence about having any object it and the type of the object § NMS is then applied to retain the most confident boxes Final detections CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 72 / 106

  55. Detection RCNN Architectures YOLO Segmentation Training YOLO § How do we train this network § Consider a cell such that a true bounding box corresponds to this cell S × S grid on input S × S grid on input § Initially the network with random weights will produce some values for these (5 + C ) values § YOLO uses sum-squared error between the predictions and the ground truth to calculate loss. The following losses are computed ◮ Classification Loss ◮ Localization Loss ◮ Confidence Loss Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 73 / 106

  56. Detection RCNN Architectures YOLO Segmentation Training YOLO Classification Loss S 2 � 2 � ✶ obj � � p i ( c ) − ˆ p i ( c ) i i =0 c ∈ classes where, ✶ obj = 1 , if a ground truth object is in cell i , otherwise 0. i p i ( c ) is the predicted probability of an object of class c in the i th cell. ˆ p i ( c ) is the ground truth label. Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 74 / 106

  57. Detection RCNN Architectures YOLO Segmentation Training YOLO Localization Loss : It measures the errors in the predicted bounding box locations and size. The loss is computed for the one box that is responsible for detecting the object. S 2 B � 2 + � 2 � � � ✶ obj �� � x i − ˆ x i − ˆ λ coord x i x i ij i =0 j =0 S 2 B �� √ w i − � � 2 + � 2 � � � ✶ obj ˆ � �� + λ coord ˆ w i h i − h i ij i =0 j =0 ij = 1 , if j th bounding box is responsible for detecting the where, ✶ obj ground truth object in cell i , otherwise 0. By square rooting the box dimensions some parity is maintained for different size boxes. Absolute errors in large boxes and small boxes are not treated same. Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 75 / 106

  58. Detection RCNN Architectures YOLO Segmentation Training YOLO Confidence Loss: For a box responsible for predicting an object S 2 B ✶ obj � 2 � � C i − ˆ � C i ij i =0 j =0 ij = 1 , if j th bounding box is responsible for detecting the where, ✶ obj ground truth object in cell i , otherwise 0. C i is the predicted probability that there is an object in the i th cell. ˆ C i is the ground truth label (of whether an object is there). Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 76 / 106

  59. Detection RCNN Architectures YOLO Segmentation Training YOLO Confidence Loss: For a box that predicts ‘no object’ inside S 2 B � 2 ✶ noobj C i − ˆ � � � λ noobj C i ij i =0 j =0 = 1 , if j th bounding box is responsible for predicting ‘no where, ✶ obj i object’ in cell i , otherwise 0. C i is the predicted probability that there is an object in the i th cell. ˆ C i is the ground truth label (of whether an object is there). The total loss is the sum of all the above losses Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 77 / 106

  60. Detection RCNN Architectures YOLO Segmentation Training YOLO Method Pascal 2007 mAP Speed DPM v5 33.7 0.07 FPS — 14 sec/ image RCNN 66.0 0.05 FPS — 20 sec/ image Fast RCNN 70.0 0.5 FPS — 2 sec/ image Faster RCNN 73.2 7 FPS — 140 msec/ image YOLO 69.0 45 FPS — 22 msec/ image CS7015 course, IIT Madras Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 78 / 106

  61. Detection RCNN Architectures YOLO Segmentation Segmentation Other Computer Vision Tasks Semantic Instance Semantic Classification Object Segmentation Segmentation Segmentation + Localization Detection GRASS , CAT , GRASS , CAT , CAT DOG , DOG , CAT DOG , DOG , CAT TREE , SKY TREE , SKY Source: cs231n course, Stanford University No objects, just pixels Abir Das (IIT Kharagpur) Multiple Object CS60010 March 04 and 05, 2020 79 / 106 No objects, just pixels Single Object This image is CC0 public domain Lecture 11 - May 10, 2018 Fei-Fei Li & Justin Johnson & Serena Yeung 8

  62. Detection RCNN Architectures YOLO Segmentation Segmentation Semantic Segmentation Idea: Sliding Window Classify center Extract patch pixel with CNN Full image Cow Cow Grass Problem: Very inefficient! Not reusing shared features between overlapping patches Farabet et al, “Learning Hierarchical Features for Scene Labeling,” TPAMI 2013 Pinheiro and Collobert, “Recurrent Convolutional Neural Networks for Scene Labeling”, ICML 2014 Lecture 11 - May 10, 2018 Fei-Fei Li & Justin Johnson & Serena Yeung 13 Source: cs231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 80 / 106

  63. Detection RCNN Architectures YOLO Segmentation Segmentation Semantic Segmentation Idea: Fully Convolutional Design a network as a bunch of convolutional layers to make predictions for pixels all at once! Conv Conv Conv Conv argmax Input: Predictions: Scores: 3 x H x W H x W C x H x W Convolutions: D x H x W Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 2018 15 Source: cs231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 81 / 106

  64. Detection RCNN Architectures YOLO Segmentation Segmentation Semantic Segmentation Idea: Fully Convolutional Design a network as a bunch of convolutional layers to make predictions for pixels all at once! Conv Conv Conv Conv argmax Input: Predictions: Scores: 3 x H x W H x W C x H x W Convolutions: Problem: convolutions at D x H x W original image resolution will be very expensive ... Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 2018 15 Source: cs231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 82 / 106

  65. Detection RCNN Architectures YOLO Segmentation Segmentation Semantic Segmentation Idea: Fully Convolutional Design network as a bunch of convolutional layers, with Downsampling : Upsampling : downsampling and upsampling inside the network! Pooling, strided ??? convolution Med-res: Med-res: D 2 x H/4 x W/4 D 2 x H/4 x W/4 Low-res: D 3 x H/4 x W/4 Input: High-res: High-res: Predictions: 3 x H x W D 1 x H/2 x W/2 D 1 x H/2 x W/2 H x W Long, Shelhamer, and Darrell, “Fully Convolutional Networks for Semantic Segmentation”, CVPR 2015 Noh et al, “Learning Deconvolution Network for Semantic Segmentation”, ICCV 2015 Lecture 11 - May 10, 2018 Fei-Fei Li & Justin Johnson & Serena Yeung 17 Source: cs231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 83 / 106

  66. Detection RCNN Architectures YOLO Segmentation Segmentation In-Network upsampling: “Unpooling” “Bed of Nails” Nearest Neighbor 1 0 2 0 1 1 2 2 1 2 0 0 0 0 1 2 1 1 2 2 3 4 3 4 3 0 4 0 3 3 4 4 0 0 0 0 3 3 4 4 Input: 2 x 2 Output: 4 x 4 Input: 2 x 2 Output: 4 x 4 Lecture 11 - Source: cs231n course, Stanford University May 10, 2018 Fei-Fei Li & Justin Johnson & Serena Yeung 18 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 84 / 106

  67. Detection RCNN Architectures YOLO Segmentation Segmentation In-Network upsampling: “Max Unpooling” Max Pooling Max Unpooling Remember which element was max! Use positions from pooling layer 0 0 2 0 1 2 6 3 1 2 … 0 1 0 0 3 5 2 1 5 6 3 4 0 0 0 0 1 2 2 1 7 8 Rest of the network 3 0 0 4 7 3 4 8 Input: 2 x 2 Output: 4 x 4 Input: 4 x 4 Output: 2 x 2 Corresponding pairs of downsampling and upsampling layers Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - 19 May 10, 2018 Source: cs231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 85 / 106

  68. Detection RCNN Architectures YOLO Segmentation Segmentation Learnable Upsampling: Transpose Convolution Recall: Normal 3 x 3 convolution, stride 1 pad 1 Dot product between filter and input Input: 4 x 4 Output: 4 x 4 Source: cs231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 86 / 106 Lecture 11 - May 10, 2018 Fei-Fei Li & Justin Johnson & Serena Yeung 22

  69. Detection RCNN Architectures YOLO Segmentation Segmentation Learnable Upsampling: Transpose Convolution Recall: Normal 3 x 3 convolution, stride 1 pad 1 Dot product between filter and input Input: 4 x 4 Output: 4 x 4 Source: cs231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 87 / 106 Lecture 11 - May 10, 2018 Fei-Fei Li & Justin Johnson & Serena Yeung 22

  70. Detection RCNN Architectures YOLO Segmentation Segmentation Learnable Upsampling: Transpose Convolution Recall: Normal 3 x 3 convolution, stride 1 pad 1 Dot product between filter and input Input: 4 x 4 Output: 4 x 4 Source: cs231n course, Stanford University Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 88 / 106 Lecture 11 - May 10, 2018 Fei-Fei Li & Justin Johnson & Serena Yeung 22

  71. Detection RCNN Architectures YOLO Segmentation Segmentation Learnable Upsampling: Transpose Convolution Recall: Normal 3 x 3 convolution, stride 2 pad 1 Dot product between filter and input Input: 4 x 4 Output: 2 x 2 Source: cs231n course, Stanford University Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - 25 May 10, 2018 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 89 / 106

  72. Detection RCNN Architectures YOLO Segmentation Segmentation Learnable Upsampling: Transpose Convolution Recall: Normal 3 x 3 convolution, stride 2 pad 1 Dot product between filter and input Input: 4 x 4 Output: 2 x 2 Source: cs231n course, Stanford University Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - 25 May 10, 2018 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 90 / 106

  73. Detection RCNN Architectures YOLO Segmentation Segmentation Learnable Upsampling: Transpose Convolution Recall: Normal 3 x 3 convolution, stride 2 pad 1 Dot product between filter and input Input: 4 x 4 Output: 2 x 2 Source: cs231n course, Stanford University Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - 25 May 10, 2018 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 91 / 106

  74. Detection RCNN Architectures YOLO Segmentation Segmentation Learnable Upsampling: Transpose Convolution 3 x 3 transpose convolution, stride 1 pad 0 Input gives weight for filter Input: 2 x 2 Output: 4 x 4 Source: cs231n course, Stanford University Lecture 11 - May 10, 2018 Fei-Fei Li & Justin Johnson & Serena Yeung 28 Abir Das (IIT Kharagpur) CS60010 March 04 and 05, 2020 92 / 106

Recommend

Semantic segmentation Image classification Object detection Semantic segmentation Evolution

Semantic segmentation Image classification Object detection Semantic segmentation Evolution

Accel : A Corrective Fusion Network for Efficient Semantic Segmentation on Video Samvit Jain , Xin Wang , Joseph Gonzalez RISE Lab, UC Berkeley Semantic segmentation Image classification Object detection Semantic segmentation Evolution

857 views • 13 slides
Segmentation and Contour Detection Image segmentation is the process of assigning a

Segmentation and Contour Detection Image segmentation is the process of assigning a

Segmentation and Contour Detection Image segmentation is the process of assigning a label to every pixel in an image such that pixels with the same label share certain visual characteristics. Contours are the lines separating

676 views • 32 slides
Detection and Segmentation CS60010: Deep Learning Abir Das IIT Kharagpur Feb 28, 2020

Detection and Segmentation CS60010: Deep Learning Abir Das IIT Kharagpur Feb 28, 2020

Detection and Segmentation CS60010: Deep Learning Abir Das IIT Kharagpur Feb 28, 2020 Introduction Datasets Localization Agenda To get introduced to two important tasks of computer vision - detection and segmentation along with deep neural

756 views • 38 slides
Rich feature hierarchies for accurate object detection and semantic segmentation Ross Girshick, Je

Rich feature hierarchies for accurate object detection and semantic segmentation Ross Girshick, Je

Rich feature hierarchies for accurate object detection and semantic segmentation Ross Girshick, Je ff Donahue, Trevor Darrell, Jitendra Malik UC Berkeley Tech Report @ http://arxiv.org/abs/1311.2524 Detection & Segmentation input

704 views • 37 slides
Introduction to Object Detection & Image Segmentation Abel Brown (abelb@nvidia.com) November

Introduction to Object Detection & Image Segmentation Abel Brown (abelb@nvidia.com) November

Introduction to Object Detection & Image Segmentation Abel Brown (abelb@nvidia.com) November 2, 2017 Outline What is Object Detection and Segmentation? Examples Before Deep Learning Common Issues with Algorithms Quality Assessment and

512 views • 47 slides
Practical Object Detection and Segmentation Vincent Chen and Edward Chou Agenda Why would

Practical Object Detection and Segmentation Vincent Chen and Edward Chou Agenda Why would

Practical Object Detection and Segmentation Vincent Chen and Edward Chou Agenda Why would understanding different architectures be useful? Modular Frameworks Describe Modern Frameworks Detection Segmentation

641 views • 38 slides
Object Detection and Segmentation from Joint Embedding of Parts and Pixels Michael Maire 1 ,

Object Detection and Segmentation from Joint Embedding of Parts and Pixels Michael Maire 1 ,

Object Detection and Segmentation from Joint Embedding of Parts and Pixels Michael Maire 1 , Stella X. Yu 2 , Pietro Perona 1 1 California Institute of Technology - Pasadena, CA 91125 2 Boston College - Chestnut Hill, MA 02467 Segmentation

740 views • 59 slides
Detection, Segmentation Overview Object Detection deer cat Object Detection as Classification

Detection, Segmentation Overview Object Detection deer cat Object Detection as Classification

CS6501: Deep Learning for Visual Recognition Detection, Segmentation Overview Object Detection deer cat Object Detection as Classification deer? cat? CNN background? Object Detection as Classification deer? cat? CNN background?

577 views • 28 slides
An Introduction to Modern Object Detection Gang Yu yugang@megvii.com Visual Recognition A

An Introduction to Modern Object Detection Gang Yu yugang@megvii.com Visual Recognition A

An Introduction to Modern Object Detection Gang Yu yugang@megvii.com Visual Recognition A fundamental task in computer vision Classification Object Detection Semantic Segmentation Instance Segmentation Key point Detection

1.13k views • 83 slides
Seminar: Medical Image Processing A robust approach for automatic detection and segmentation of

Seminar: Medical Image Processing A robust approach for automatic detection and segmentation of

Introduction Morphology Linear filters Detection Evaluation Summary Seminar: Medical Image Processing A robust approach for automatic detection and segmentation of cracks in underground pipeline images Tim Niemueller < tim@niemueller.de

2.1k views • 165 slides
Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation Authors: Ross

Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation Authors: Ross

Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation Authors: Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra Malik Presented by Huihuang Zheng Problem: Object Detection Regionlets SegDPM (2013) Selective

499 views • 19 slides
Detection and Segmentation of Detection and Segmentation of Touching Characters in Touching

Detection and Segmentation of Detection and Segmentation of Touching Characters in Touching

Detection and Segmentation of Detection and Segmentation of Touching Characters in Touching Characters in Mathematical Expressions Mathematical Expressions A. Nomura, A. Nomura, K. Michishita, K. Michishita, S. Uchida, S. Uchida, M.

571 views • 19 slides
LSB Galaxies Detection Using Markovian Segmentation on Astronomical Images Mireille Louys 1 ,

LSB Galaxies Detection Using Markovian Segmentation on Astronomical Images Mireille Louys 1 ,

LSB Galaxies Detection Using Markovian Segmentation on Astronomical Images Mireille Louys 1 , Benjamin Perret 1 Bernd Vollmer 2 , Franois Bonnarel 2 Sebastien Lefvre 1 , Christophe Collet 1 1 L aboratoire des S ciences de l I nformatique ,

318 views • 13 slides
Contour Detection and Hierarchical Image Segmentation P. Arbelaez, M. Maire, C. Fowlkes, and J.

Contour Detection and Hierarchical Image Segmentation P. Arbelaez, M. Maire, C. Fowlkes, and J.

Contour Detection and Hierarchical Image Segmentation P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik. IEEE TPAMI 2011. Islam Beltagy Original slides from: Hsin-Min Cheng http://archer.ee.nctu.edu.tw/powerpoint/GM_1024.pptx Outline

644 views • 31 slides
Making Object Detection Work in Medical Imaging Idan Bassuk Segmentation Detection

Making Object Detection Work in Medical Imaging Idan Bassuk Segmentation Detection

Making Object Detection Work in Medical Imaging Idan Bassuk Segmentation Detection Classification 4 5 Mature technology 6 Mature, Accurate and Fast 7 Changing the world, one killer app at a time 8 Changing the world, one killer app at

570 views • 39 slides
Segmentation Driven Object Detection with Introduction Fisher Vectors State of the art Method

Segmentation Driven Object Detection with Introduction Fisher Vectors State of the art Method

1/21 Segmentation Driven Object Detection with Fisher Vectors Camille Brasseur Segmentation Driven Object Detection with Introduction Fisher Vectors State of the art Method Evaluation Camille Brasseur Conclusions 20 dcembre 2013

490 views • 22 slides
Modern Object Detection Gang Yu Face++ Researcher yugang@megvii.com Visual Recognition A

Modern Object Detection Gang Yu Face++ Researcher yugang@megvii.com Visual Recognition A

Lecture 6: Modern Object Detection Gang Yu Face++ Researcher yugang@megvii.com Visual Recognition A fundamental task in computer vision Classification Object Detection Semantic Segmentation Instance Segmentation Key point

1.24k views • 85 slides
speaker change detection using fundamental frequency with application to multi-talker segmentation

speaker change detection using fundamental frequency with application to multi-talker segmentation

speaker change detection using fundamental frequency with application to multi-talker segmentation May 16, 2019 Aidan Hogg, Christine Evers and Patrick Naylor Electrical and Electronic Engineering, Imperial College London, UK diarization

735 views • 29 slides
External and Intrinsic Plagiarism Detection using a Cross-Lingual Retrieval and Segmentation

External and Intrinsic Plagiarism Detection using a Cross-Lingual Retrieval and Segmentation

External and Intrinsic Plagiarism Detection using a Cross-Lingual Retrieval and Segmentation System Markus Muhr, Roman Kern, Mario Zechner, Michael Granitzer { mmuhr, rkern, mzechner, mgrani } @know-center.at CLEF 2010 / PAN / 2010-09-22

282 views • 15 slides
Comparing Objective Functions for Segmentation and Detection of Microaneurysms in Retinal Images

Comparing Objective Functions for Segmentation and Detection of Microaneurysms in Retinal Images

Comparing Objective Functions for Segmentation and Detection of Microaneurysms in Retinal Images Medical Imaging with Deep Learning, Montral, 6 9 July 2020 Jakob K. H. Andersen 1,3,* Jakob Grauslund 2,3 Thiusius R. Savarimuthu 1 1 The Maersk

561 views • 11 slides
Other Related Problems Applications Classification Instance Classification Object Detection +

Other Related Problems Applications Classification Instance Classification Object Detection +

Image Segmentation Goal: Label every pixel with its semantic category * Other Related Problems Applications Classification Instance Classification Object Detection + Localization Segmentation Background Subtraction Cut-and-paste

642 views • 23 slides
Change: Detection, Estimation, Segmentation Abstract The maximum score statistic is used to

Change: Detection, Estimation, Segmentation Abstract The maximum score statistic is used to

Change: Detection, Estimation, Segmentation Abstract The maximum score statistic is used to detect and estimate changes in the level, slope, or other local feature of a sequence of observations, and to segment the sequence when there appear to

784 views • 19 slides
VIDEO SIGNALS Segmentation WHAT IS SEGMENTATION WHAT IS SEGMENTATION Segmentation is a

VIDEO SIGNALS Segmentation WHAT IS SEGMENTATION WHAT IS SEGMENTATION Segmentation is a

VIDEO SIGNALS Segmentation WHAT IS SEGMENTATION WHAT IS SEGMENTATION Segmentation is a fundamental low level operation on Segmentation is a fundamental low-level operation on images. If an image is already partitioned into segments

639 views • 51 slides
Real-Time Object Detection and Semantic Segmentation Geetank Raipuria Wei Li Corporate

Real-Time Object Detection and Semantic Segmentation Geetank Raipuria Wei Li Corporate

Real-Time Object Detection and Semantic Segmentation Geetank Raipuria Wei Li Corporate Development manager Computer Vision Engineer NavInfo Europe Presented at GTC 2019 Session S9351 NavInfo - our growth benefits from AI Smart Mobility

735 views • 34 slides

More recommend