Research map / Detection and segmentation / Open-set and few-shot detection
Learning Dense Object Descriptors from Multiple Views for Low-shot Category Generalization
Stefan Stojanov, Ngoc Anh Thai, Zixuan Huang, James Matthew Rehg
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 |
|---|---|---|---|
| Train on ABC using local patch correspondences rather than global instance-level features | DOPE versus VISPE++ - MoCoV2 · ModelNet40-LS, 5-way 1-shot | accuracy: 61.07 \ci{0.41} -> 62.76 \ci{0.43} | Improved |
| Train on ABC using local patch correspondences rather than global instance-level features | DOPE versus VISPE++ - SimSiam · ModelNet40-LS, 5-way 5-shot | accuracy: 76.55 \ci{0.22} -> 76.86 \ci{0.33} | Improved |
| Train on ABC using local patch correspondences rather than global instance-level features | DOPE versus VISPE++ - MoCoV2 · ShapeNet-LS, 5-way 1-shot | accuracy: 58.99 \ci{0.46} -> 62.00 \ci{0.48} | Improved |
| Train on ABC using local patch correspondences rather than global instance-level features | DOPE versus VISPE++ - MoCoV2 · ShapeNet-LS, 5-way 5-shot | accuracy: 72.26 \ci{0.39} -> 73.55 \ci{0.38} | Improved |
| Use local patch matching rather than supervised FEAT | DOPE versus FEAT · CO3D-LS, 5-way 1-shot | accuracy: 64.48 \ci{0.43} -> 61.77 \ci{0.44} | Hurt |
| Use local patch matching rather than supervised RFS | DOPE versus RFS · CO3D-LS, 5-way 5-shot | accuracy: 81.50 \ci{0.26} -> 75.16 \ci{0.32} | Hurt |
| Include negatives from other object instances as well as the second view (ablation) | DOPE · ModelNet40-LS, 5-way 1-shot | accuracy: 61.51 \ci{0.42} -> 62.76 \ci{0.43} | Improved |
| Include negatives from the second view as well as other objects (ablation) | DOPE · ModelNet40-LS, 5-way 1-shot | accuracy: 43.50 \ci{0.39} -> 62.76 \ci{0.43} | Improved |
| Train with physically distinct object viewpoints rather than two augmentations of one image (ablation) | DOPE · ModelNet40-LS, 5-way 1-shot | accuracy: 38.36 \ci{0.43} -> 62.76 \ci{0.43} | Improved |
| Add random background removal during training (ablation) | DOPE · ModelNet validation classes, 1-shot 5-way | accuracy: 54.24 -> 61.78 | Improved |
| Add mask prediction and mask local feature encodings (ablation) | DOPE · ModelNet validation classes, 1-shot 5-way | accuracy: 50.06 -> 61.78 | Improved |
| Replace known foreground masks with FreeSOLO-predicted masks (ablation) | DOPE - ABC · ModelNet40-LS, 5-way 1-shot | accuracy: 62.76 \ci{0.43} -> 57.63 \ci{0.40} | Hurt |
Closest papers by meaning
- Self-Supervised Learning of Object Parts for Semantic Segmentation (CVPR 2022 · similarity 0.87)
- Large Scale Visual Recognition through Adaptation using Joint Representation and Multiple Instance Learning (JMLR 2016 · similarity 0.86)
- Unsupervised Object-Level Representation Learning from Scene Images (NeurIPS 2021 · similarity 0.86)
- Rethinking Generalization in Few-Shot Classification (NeurIPS 2022 · similarity 0.86)
- Learning from Small Sample Sets by Combining Unsupervised Meta-Training with CNNs (NeurIPS 2016 · similarity 0.86)
- Near, far: Patch-ordering enhances vision foundation models' scene understanding (ICLR 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.
