Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Why Can't I Open My Drawer? Mitigating Object-Driven Shortcuts in Zero-Shot Compositional Action Recognition
Geo Ahn, Inwoong Lee, Taeoh Kim, Minho Shim, Dongyoon Wee, Jinwoo Choi
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 CPR and TORC to C2C (ablation) | C2C · Sth-com val (CLIP) | unseen composition accuracy: 30.14 -> 35.16 | Improved |
| add TORC alone (ablation) | C2C · Sth-com val (CLIP) | verb@unseen-comp: 54.02 -> 57.49 | Improved |
| add CPR alone (ablation) | C2C · Sth-com val (CLIP) | unseen vs seen composition accuracy: unseen 30.14 -> 33.22; seen 47.31 -> 43.09 | Mixed |
| add L_cos temporal-reversal term (ablation) | TORC · Sth-com val (CLIP) | unseen composition accuracy: 30.14 -> 31.11 | Improved |
| add entropy term on shuffled features (ablation) | TORC · Sth-com val (CLIP) | unseen composition accuracy: 30.14 -> 30.84 | Improved |
| use both L_cos and L_ent (ablation) | TORC · Sth-com val (CLIP) | unseen composition accuracy: 30.14 -> 33.92 | Improved |
| optimize cross-entropy over the full verb-object space instead of batch-adaptive (ablation) | CPR · Sth-com val (CLIP) | unseen composition accuracy: 33.05 -> 17.34 | Hurt |
| use batch-adaptive label expansion instead of closed-world labels (ablation) | CPR · Sth-com val (CLIP) | unseen composition accuracy: 30.80 -> 33.05 | Improved |
| penalize all hard negatives instead of only frequent ones (ablation) | CPR · Sth-com val (CLIP) | composition H.M.: 37.67 -> 31.33 | Hurt |
| penalize frequent hard negatives with L_CPR (ablation) | CPR · Sth-com val (CLIP) | composition H.M.: 37.28 -> 37.67 | Improved |
| apply RCORE vs C2C on Sth-com | C2C · Sth-com test (CLIP) | unseen composition accuracy: 30.08 -> 33.90 | Improved |
| apply RCORE vs C2C on EK100-com | C2C · EK100-com test (CLIP) | unseen composition accuracy: 21.56 -> 28.41 | Improved |
Closest papers by meaning
- C2C: Component-to-Composition Learning for Zero-Shot Compositional Action Recognition (ECCV 2024 · similarity 0.90)
- Compositional Learning for Human Object Interaction (ECCV 2018 · similarity 0.89)
- Beyond Seen Primitive Concepts and Attribute-Object Compositional Learning (CVPR 2024 · similarity 0.87)
- ProCC: Progressive Cross-Primitive Compatibility for Open-World Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.85)
- Troika: Multi-Path Cross-Modal Traction for Compositional Zero-Shot Learning (CVPR 2024 · similarity 0.85)
- Kronecker Mask and Interpretive Prompts are Language-Action Video Learners (ICLR 2025 · similarity 0.84)
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.
