Research map / Learning theory and algorithms / Prediction and decision losses
Learning Generalized Intersection Over Union for Dense Pixelwise Prediction
Jiaqian Yu, Jingtao Xu, Yiwei Chen, Weiming Li, Qiang Wang, Byungin Yoo, Jae-Joon Han
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Lovász IoU loss with Lovász PixIoU loss | AFOD pixelwise object tracking · VOT-ST2020 | mIoU(%): +1.4 over Lv. IoU (74.2 to 75.6); +4.0 over CE | Improved |
| Replace Lovász IoU loss with Lovász PixIoU loss | DeepLabv3 ResNet-50 · Pascal VOC 2012 val | mIoU(%): +0.3 over Lv. IoU (80.5 to 80.8) | Improved |
| Replace Lovász IoU loss with Lovász PixIoU loss | DeepLabv3 ResNet-101 · Pascal VOC 2012 val | mIoU(%): +0.3 over Lv. IoU (82.8 to 83.1); +0.2 over CE | Improved |
| Replace Lovász IoU loss with Lovász PixIoU loss | DeepLabV3+ ResNet-101 with pseudo-pretrain · Cityscapes val | mIoU(%): +0.82 over Lv. IoU (79.80 to 80.62); +0.64 over WCE | Improved |
| Replace Lovász IoU loss with Lovász PixIoU loss | DeepLabV3+ ResNet-101 with pseudo-pretrain · Cityscapes val | iIoU: +0.65 over Lv. IoU (62.54 to 63.19); +0.88 over WCE | Improved |
| Replace Lovász IoU loss with Lovász PixIoU loss | Panoptic-DeepLab ResNet-50 · Cityscapes val | mIoU(%): +0.56 over Lv. IoU (78.45 to 79.01); +0.29 over WCE | Improved |
| Replace Lovász IoU loss with Lovász PixIoU loss | Panoptic-DeepLab ResNet-50 · Cityscapes val | PQ: 60.26 vs 60.28 for WCE; vs 59.47 for Lv. IoU | No clear effect |
| Replace cross-entropy with Lovász-based losses (IoU or PixIoU) on DeepLabv3 ResNet-101 | DeepLabv3 ResNet-101 · Pascal VOC 2012 val | mIoU(%): Lv. IoU 82.8 vs CE 82.9; Lv. Pix 83.1 | Mixed |
| Use the Lovász PixIoU loss on Cityscapes WCE per-class mIoU | DeepLabV3+ ResNet-101 · Cityscapes val | per-class mIoU (wall): wall 57.05 (WCE) to 62.88 (Lv. Pix), +10.21% | Improved |
| Replace Lovász IoU loss with Lovász PixIoU loss in synthetic linear model training (ablation) | 3x3 conv linear model on synthetic data · Synthetic triangle shapes | mean PixIoU and IoU: qualitative/curve comparison; no numeric values given | Improved |
| Train from scratch with Lovász IoU instead of pseudo-pretrain strategy (ablation) | DeepLabV3+ ResNet-101 · Cityscapes val | mIoU: 76.64 mIoU from scratch vs 79.8 with pseudo-pretrain | Hurt |
| Pseudo-pretrain strategy (5k WCE iterations before loss switch) versus WCE only for same total iterations (ablation) | DeepLabV3+ ResNet-101 · Cityscapes val | mIoU: WCE with pseudo-pretrain 79.98 vs released 79.98 | No clear effect |
Closest papers by meaning
- Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression (CVPR 2019 · similarity 0.90)
- Deep Dive Into Gradients: Better Optimization for 3D Object Detection With Gradient-Corrected IoU Supervision (CVPR 2023 · similarity 0.86)
- $\alpha$-IoU: A Family of Power Intersection over Union Losses for Bounding Box Regression (NeurIPS 2021 · similarity 0.85)
- Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression (AAAI 2020 · similarity 0.85)
- RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation (NeurIPS 2025 · similarity 0.79)
- Rethinking Boundary Discontinuity Problem for Oriented Object Detection (CVPR 2024 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
