hang zhao massachusetts institute of technology last year
play

Hang Zhao Massachusetts Institute of Technology Last years - PowerPoint PPT Presentation

Hang Zhao Massachusetts Institute of Technology Last years challenge at ECCV ADE Dataset New tasks this year Instance Segmentation Instance Segmentation 2.5 10 4 # instances Sta$s$cs: 2 100 foreground categories 200k


  1. Hang Zhao Massachusetts Institute of Technology

  2. Last year’s challenge at ECCV

  3. ADE Dataset

  4. New tasks this year …

  5. Instance Segmentation

  6. Instance Segmentation 2.5 × 10 4 # instances • Sta$s$cs: 2 • 100 foreground categories • 200k instances 1.5 • #Instances per category: 200 ~ 20k • #Instances per image: ~11 1 • Evalua$on: • Mean AP over 100 categories 0.5 0 10 20 30 40 50 60 70 80 90 100

  7. Semantic Boundary Detection Perceptual Edges

  8. Semantic Boundary Detection Binary Edges Category-aware Semantic Boundaries

  9. Semantic Boundary Detection Boundary between: person and chair Boundary between: person and sofa Boundaries reveal object interac8ons

  10. Semantic Boundary Detection • 150 categories • Inter-class boundaries • Intra-class boundaries • Evalua$on: • F-measure at op$mal dataset scale (F-ODS)

  11. Results – Scene Parsing Final Score = (mean IoU + pixel accuracy) / 2 CASIA_IVA_JD Jun Fu, Jing Liu, Longteng Guo, Haijie Team Name Final Score Tian, Fei Liu, Yong Li, Yongjun Bao, CASIA_IVA_JD 0.5547 Weipeng Yan, Hanqing Lu CASIA IVA and JD.com WinterIsComing 0.5544 xdliang 0.5447 lijc15 0.5294 WinterIsComing … … Riwei Chen, Qi Chen, Xinglong Wu, Yifan Lu, Yudong Jiang, Linfu Wen baseline-DilatedNet 0.4567 Tou8ao AI Lab, ByteDance * We received 30 submissions from 10 teams in this task. Full ranking can be found on places challenge website.

  12. Sample Results Scene Parsing

  13. Image Ground Truth Predic$on

  14. Image Ground Truth Predic$on

  15. Results – Instance Segmentation Megvii (Face++) Team Name mean AP * Tete Xiao, * Ruixuan Luo, * Borui Jiang, Shuai Shao, Yuning Jiang, Yadong Mu, Megvii (Face++) 0.2977 Jieqi Shi, Chi Zhang, Jian Sun Megvii Research and Peking University G-RMI 0.2415 BlueSky 0.1551 … … G-RMI Alireza Fathi, Nori Kanazawa, Kai Yang, baseline 0.200 Kevin Murphy Google Research Machine Intelligence * We received 15 submissions from 5 teams in this task. Full ranking can be found on places challenge website.

  16. Sample Results Instance Segmentation

  17. Results – Semantic Boundary Detection State-of-the-art: CASENet CVPR’17 hmp://www.merl.com/research/license#CASENet Z. Yu, et al. CASENet: Deep category-aware semantic edge detection

  18. Sample Results Semantic Boundary Detection

  19. Image Ground Truth Predic$on

  20. Image Ground Truth Predic$on

  21. Awards Winner of Scene Parsing: CASIA_IVA_JD Jun Fu, Jing Liu, Longteng Guo, Haijie Tian, Fei Liu, Yong Li, Yongjun Bao, Weipeng Yan, Hanqing Lu CASIA IVA and JD.com Prize: Titan Xp GPU Sponsored by: NVIDIA

  22. Awards Winner of Instance Segmenta8on: Megvii (Face++) * Tete Xiao, * Ruixuan Luo, * Borui Jiang, Shuai Shao, Yuning Jiang, Yadong Mu, Jieqi Shi, Chi Zhang, Jian Sun Megvii Research and Peking University Prize: Titan Xp GPU Sponsored by: NVIDIA

Recommend


More recommend