Experience on CrowdHuman Challenge Zheng Ge 1 2
- Slides: 13
Experience on Crowd-Human Challenge Zheng Ge 1, 2 Xin Huang 1, 2 Zequn Jie 1, * Yuhu Shan 1 1 2 Waseda University Tencent AI Lab * Team leader
Baseline Re-implementation Exploring Techniques for Further Improvement
Baseline Re-implementation Baseline (visible body, paper): Faster-RCNN, Res 50, FPN m. MR: 55. 94% m. MR: 59. 67% Baseline (visible body, ours): Faster-RCNN, Res 50, FPN, ROI_align FPN+BN, Avoiding negative anchors in ignored regions m. MR: 55. 76%
Baseline Re-implementation ignore region anchor An example of labeled ignored region. During training RPN, we avoid the negative anchors whose Io. A (Intersection over Anchor) with an arbitrary labeled ignored region > 0. 5.
Baseline Re-implementation m. MR AP@0. 5 Baseline (paper) 55. 94 85. 60 Baseline (ours) 59. 67 83. 70 + Avoid negative anchors 58. 53 84. 92 + BN FPN 55. 76 85. 43 Table 1. Evaluation results on Crowdhuman (visible body) validation set. Baseline (ours): Faster RCNN, Res 50, FPN, ROI_align, nms_pre=6000
Baseline Re-implementation The effect of nms_pre m. MR AP@0. 5 12000 56. 45 85. 37 6000 55. 76 85. 43 2000 54. 24 84. 00 1500 54. 00 83. 38 1000 53. 86 82. 03 Table 2. Evaluation results on Crowdhuman (visible body) validation set. The best result on previous page
Baseline Re-implementation Conclusion m. MR AP@0. 5 Baseline (paper) 55. 94 85. 60 Baseline (ours new) 54. 24 84. 00 Table 3. Baseline comparison on validation set (visible body). m. MR AP@0. 5 Baseline (paper) 50. 42 84. 95 Baseline (ours new) 46. 52 84. 04 Table 4. Baseline comparison on validation set (full body). Baseline (ours new): Faster RCNN, Res 50, BN FPN, ROI_align, nms_pre=2000, Avoid negative anchors
Baseline Re-implementation Exploring Techniques for Further Improvement
Exploring Techniques bringing large improvements • • • Cascade R-CNN Deformable Conv Net SENet 154 Multi-Scale Train/Test Ensemble Techniques bringing marginal improvements • Bounded Repulsion Loss • R-CNN Context • Merge Cityperson dataset Techniques of limited effects • • • Sync. BN Focal Loss GIo. U/Io. U Loss COCO Pretrain Soft-NMS OHEM Adaptive-NMS Guided Anchor RPN Scale Balanced Sampling …
Exploring Techniques Conclusion m. MR gain Baseline 46. 52 - + Bounded Rep. GT Loss 45. 80 0. 72 + R-CNN Context 45. 45 0. 35 + multi-scale train/test 43. 32 2. 13 + Cascade R-CNN - - + Deformable Conv - - 37. 17 6. 15 + SENet 154 m. MR gain Baseline 46. 52 - + City. Person 45. 58 0. 94 Baseline : Faster RCNN, Res 50, BN FPN, ROI_align, nms_pre=2000, Avoid anchors
Exploring Techniques Final Ensemble for Submission • • • SENet 154 + Cascade R-CNN + DCN + Bounded Rep. Gt + Context + cocopretrain SENet 154 + Cascade R-CNN + DCN + Bounded Rep. Gt + Context + cityperson SERes. Ne. Xt 101 + Cascade R-CNN + DCN + Bounded Rep. Gt + Context
Exploring Techniques About Jaccard Index Score Threshold Val. Set filtering Val. Subset best threshold: 0. 55 huge gap Test Set Thr. JI 0. 55 75. 78 0. 50 76. 48 0. 45 77. 01 0. 40 77. 46
Thanks
