Research map / Label-efficient and robust learning / Self-supervised and contrastive
Improving Equivariance in State-of-the-Art Supervised Depth and Normal Predictors
Yuanyi Zhong, Anand Bhattad, Yu-Xiong Wang, David Forsyth
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 equivariant loss on top of augmentation | U-Net · Taskonomy Depth-ZBuffer (supervised, U-Net) | delta>1.25 (%): 27.0 -> 25.0 | Improved |
| add equivariant loss on top of augmentation | U-Net · Taskonomy Edge2D | L1 error x10^-3: 7.35 -> 6.35 | Improved |
| add equivariant loss on top of augmentation | U-Net · Taskonomy Surface Normal | angular error (deg): 6.55 -> 6.47 | Improved |
| add unlabeled data with equivariant loss (ablation) | U-Net, 1/4 labeled · Taskonomy Depth-ZBuffer | delta>1.25 (%): 42.0 -> 41.0 | Improved |
| add unlabeled data with equivariant loss | U-Net, all labeled · Taskonomy Depth-ZBuffer | delta>1.25 (%): 25.0 -> 25.0 | No clear effect |
| unsupervised equivariant finetuning | MiDaS-2.1 CNN · NYU-v2 | delta>1.25 (%): 8.71 -> 7.82 | Improved |
| unsupervised equivariant finetuning | MiDaS-3.0 DPT · NYU-v2 | delta>1.25 (%): 8.32 -> 7.75 | Improved |
| unsupervised equivariant finetuning | Bae et al. normal · NYU-v2 | mean angular error (deg): 16.03 -> 15.71 | Improved |
| add inference-time 3-crop averaging (ablation) | MiDaS-v2.1 pre-trained · NYU-v2 | delta>1.25 (%): 8.71 -> 8.63 | No clear effect |
| use feature space (L-1) instead of label space (L) (ablation) | EqLoss · Taskonomy Depth-ZBuffer (supervised, U-Net) | delta>1.25 (%): 25.7 -> 25.0 | Improved |
| apply loss at earlier upsampling block up3 instead of L-1 (ablation) | EqLoss · Taskonomy Depth-ZBuffer (supervised, U-Net) | delta>1.25 (%): 25.0 -> 26.1 | Hurt |
| raise loss coefficient from 1e-4 to 3e-4 (ablation) | EqLoss · Taskonomy Depth-ZBuffer (supervised, U-Net) | delta>1.25 (%): 25.0 -> 25.8 | Hurt |
Closest papers by meaning
- Soft Equivariance Regularization for Invariant Self-Supervised Learning (ICLR 2026 · similarity 0.88)
- Normalization-Equivariant Neural Networks with Application to Image Denoising (NeurIPS 2023 · similarity 0.88)
- Tunable Soft Equivariance with Guarantees (CVPR 2026 · similarity 0.88)
- A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs (ICML 2024 · similarity 0.86)
- A Regularization-Guided Equivariant Approach for Image Restoration (CVPR 2025 · similarity 0.85)
- Rotation-Equivariant Self-Supervised Method in Image Denoising (CVPR 2025 · 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.
