Research map / Label-efficient and robust learning / Recognition and long-tail
Unsupervised Learning of Optical Flow with Deep Feature Similarity
Woobin Im, Tae-Kyun Kim, Sung-Eui Yoon
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 occlusion handling to RGB photometric loss (ablation) | PWC-Net with feature separation loss · FlyingChairs test | EPE: 4.01 -> 3.64 | Improved |
| use census transform instead of RGB features (ablation) | PWC-Net with feature separation loss · FlyingChairs test | EPE: 3.64 -> 3.10 | Improved |
| add data distillation Ld to census loss (ablation) | PWC-Net with feature separation loss · FlyingChairs test | EPE: 3.10 -> 2.93 | Improved |
| add feature separation loss Lf and regulated census loss Lr on top of Ld (ablation) | PWC-Net with feature separation loss · FlyingChairs test | EPE: 2.93 -> 2.69 | Improved |
| add feature separation loss Lf and regulated census loss Lr on top of Ld (ablation) | PWC-Net with feature separation loss · Sintel Final ALL (trained on Sintel) | EPE: 3.86 -> 3.57 | Improved |
| add Lf without Lr (ablation) | PWC-Net with feature separation loss · Sintel | EPE | No clear effect |
| add regulated census loss Lr to the loss (ablation) | PWC-Net with feature separation loss · Sintel / FlyingChairs | EPE | Improved |
| learn deep features with triplet loss instead of proposed losses (ablation) | PWC-Net with feature separation loss · Sintel / FlyingChairs | EPE | No clear effect |
| use large smoothness weight 10^-2 instead of 10^-4 (ablation) | PWC-Net with feature separation loss · FlyingChairs | EPE: 3.27 vs. 2.85 | Hurt |
| raise separation weight lambda_f to 5.0 from 4.0 (ablation) | PWC-Net with feature separation loss · FlyingChairs | EPE: 3.21 vs. 2.85 | Hurt |
| add Lf and Lr for large displacements 40+ px (ablation) | PWC-Net with feature separation loss · Sintel Final (trained on Sintel) | EPE: 27.17 -> 22.36 | Improved |
| fine-tune on KITTI with proposed losses vs DDFlow | PWC-Net with feature separation loss · KITTI 2015 test | Fl: 14.29% -> 13.38% | Improved |
Closest papers by meaning
- UnFlow: Unsupervised Learning of Optical Flow With a Bidirectional Census Loss (AAAI 2018 · similarity 0.90)
- Unsupervised Deep Learning for Optical Flow Estimation (AAAI 2017 · similarity 0.90)
- Learning by Analogy: Reliable Supervision From Transformations for Unsupervised Optical Flow Estimation (CVPR 2020 · similarity 0.89)
- What Matters in Unsupervised Optical Flow (ECCV 2020 · similarity 0.88)
- UnSAMFlow: Unsupervised Optical Flow Guided by Segment Anything Model (CVPR 2024 · similarity 0.87)
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.85)
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.
