Research map / Video / Video segmentation and tracking
Rethinking Image-to-Video Adaptation: An Object-centric Perspective
Rui Qian, Shuangrui Ding, Dahua Lin
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 object distillation loss L_obj (ablation) | Object-centric image-to-video adaptation · Kinetics-400 | Top-1 accuracy: 70.4 -> 80.3 | Improved |
| add temporal reasoning loss L_temp (ablation) | Object-centric image-to-video adaptation · Something-Something-v2 | Top-1 accuracy: 44.8 -> 52.1 | Improved |
| add both L_obj and L_temp (ablation) | Object-centric image-to-video adaptation · Something-Something-v2 | Top-1 accuracy: 44.8 -> 71.6 | Improved |
| add L_temp (ablation) | Object-centric image-to-video adaptation · DAVIS-2017-UVOS | J&F: 27.3 -> 30.5 | Improved |
| compress frames into 8 object tokens instead of keeping all original HW features (ablation) | Temporal modeling input · Something-Something-v2 | Top-1 accuracy: 64.3 -> 71.2 | Improved |
| increase object tokens from 8 to 16 (ablation) | Slot attention · Kinetics-400 / SSv2 | Top-1 accuracy: 85.8 -> 85.9 (K-400); 71.2 -> 71.2 (SSv2) | No clear effect |
| increase object tokens from 1 to 8 (ablation) | Slot attention · Something-Something-v2 | Top-1 accuracy: 44.1 -> 71.2 | Improved |
| add temporal attention module on top of object-time interaction (ablation) | Object-time interaction · Something-Something-v2 | Top-1 accuracy: 71.6 -> 66.8 | Hurt |
| add object state change modeling (Phi_ti) to object interaction module (ablation) | Object-time interaction · Something-Something-v2 | Top-1 accuracy: 50.0 -> 71.6 | Improved |
| vary sampling interval delta from 2 to larger values (or all intervals) (ablation) | State change modeling · Something-Something-v2 | Top-1 accuracy: 71.2 (delta=2) vs 71.2 (All) | No clear effect |
| apply CRF refinement to zero-shot segmentation masks | Ours K-400 · DAVIS-2017-UVOS | J&F: 44.7 -> 49.3 | Improved |
| train object tokens on K-400 rather than SSv2 (ablation) | Zero-shot video object segmentation · DAVIS-2017-UVOS | J&F: 41.0 -> 44.7 | Improved |
Closest papers by meaning
- Temporally Consistent Object-Centric Learning by Contrasting Slots (CVPR 2025 · similarity 0.89)
- Object-Centric Learning for Real-World Videos by Predicting Temporal Feature Similarities (NeurIPS 2023 · similarity 0.89)
- Rethinking Temporal Consistency in Video Object-Centric Learning: From Prediction to Correspondence (ICML 2026 · similarity 0.89)
- Reconstruction-Guided Slot Curriculum: Addressing Object Over-Fragmentation in Video Object-Centric Learning (CVPR 2026 · similarity 0.89)
- AIM: Adapting Image Models for Efficient Video Action Recognition (ICLR 2023 · similarity 0.88)
- Where Do We (Not) Need Temporal Context in Low-Resource Video Task Adaptation? (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.
