Research map / Detection and segmentation / Open-set and few-shot detection
Self-Supervised Representation Learning From Flow Equivariance
Yuwen Xiong, Mengye Ren, Wenyuan Zeng, Raquel Urtasun
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add optical flow matching to pixel-based equivariance (ablation) | FlowE · UrbanCity ablation (16K iters) | mIoU: 28.7 -> 37.9 | Improved |
| add random affine transformation (ablation) | FlowE (pixel + flow) · UrbanCity ablation (16K iters) | mIoU: 37.3 -> 37.9 | Improved |
| add random affine transformation (ablation) | FlowE (pixel + flow) · UrbanCity ablation (16K iters), heavier header | mIoU: 51.9 -> 53.2 | Improved |
| keep spatial pixel-based representation instead of global average pooling (ablation) | FlowE · UrbanCity ablation (16K iters) | mIoU: 17.8 -> 37.9 | Improved |
| add pixel-based loss to affine alone (ablation) | FlowE (affine + flow) · UrbanCity ablation (16K iters) | mAP: 0.7 -> 3.8 | Improved |
| use RAFT instead of weaker PWC-Net flow (ablation) | FlowE · UrbanCity ablation (16K iters) | mIoU: 37.2 -> 37.9 | No clear effect |
| use FlowE pretraining instead of ImageNet supervised | ResNet-50 · UrbanCity (frozen readout) | mIoU: 39.6 -> 49.6 | Improved |
| use FlowE pretraining instead of ImageNet BYOL, heavier header | ResNet-50 · BDD100K (heavier header) | mIoU: 52.4 -> 49.8 | Hurt |
| add FlowE pretraining before finetuning with 1% labels | ResNet-50 · UrbanCity semantic segmentation | mIoU: 42.0 -> 53.9 | Improved |
| add FlowE pretraining before finetuning with 100% labels | ResNet-50 · UrbanCity semantic segmentation | mIoU: 63.3 -> 68.8 | Improved |
| apply BYOL with video frames and pre-crop (ablation) | BYOL · UrbanCity (heavier header) | mIoU: 26.6 -> 13.0 | Hurt |
| use FlowE instead of BYOL on driving video | BYOL · UrbanCity (heavier header) | mIoU: 19.6 -> 61.7 | Improved |
Closest papers by meaning
- Self-Supervised Learning of Video-Induced Visual Invariances (CVPR 2020 · similarity 0.88)
- Self-Supervised Learning Based on Transformed Image Reconstruction for Equivariance-Coherent Feature Representation (AAAI 2026 · similarity 0.88)
- PooDLe🐩: Pooled and dense self-supervised learning from naturalistic videos (ICLR 2025 · similarity 0.88)
- FlowFeat: Pixel-Dense Embedding of Motion Profiles (NeurIPS 2025 · similarity 0.88)
- In-Context Symmetries: Self-Supervised Learning through Contextual World Models (NeurIPS 2024 · similarity 0.87)
- H-Flow: Self-supervised Human Scene Flow via Physics-inspired Joint Multi-modal Learning (NeurIPS 2026 · similarity 0.87)
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.
