Research map / Detection and segmentation / Open-set and few-shot detection
Does Visual Pretraining Help End-to-End Reasoning?
Chen Sun, Calvin Luo, Xingyi Zhou, Anurag Arnab, Cordelia Schmid
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 |
|---|---|---|---|
| Replace supervised visual pretraining with IV-CL self-supervised video pretraining | ViT-B visual encoder · CATER | accuracy (%): 56.64% -> 70.14 (±0.59)% | Improved |
| Replace supervised visual pretraining with IV-CL self-supervised video pretraining | ViT-B visual encoder · ACRE compositionality | accuracy (%): 67.27% -> 93.27 (±0.22)% | Improved |
| Use object detection rather than classification pretraining | ViT-B visual encoder · CATER | accuracy (%): 41.48% -> 56.64% | Improved |
| Use larger-scale JFT classification pretraining instead of ImageNet-21k classification pretraining | ViT-B visual encoder · ACRE compositionality | accuracy (%): 60.73% -> 48.32% | Hurt |
| Use IV-CL instead of video MAE pretraining | End-to-end visual reasoning model · CATER static | Top-1 accuracy (%): 63.7 -> 70.1 ± 0.6 | Improved |
| Use IV-CL instead of image MAE pretraining | End-to-end visual reasoning model · ACRE compositionality | accuracy (%): 80.27 -> 93.27 ± 0.22 | Improved |
| Use 37.5% rather than 12.5% masking (ablation) | IV-CL · CATER | accuracy (%): 66.35% -> 70.14% | Improved |
| Use 37.5% rather than 50% masking (ablation) | IV-CL · CATER | accuracy (%): 66.57% -> 70.14% | Improved |
| Use 37.5% rather than 87.5% masking (ablation) | IV-CL · CATER | accuracy (%): 61.94% -> 70.14% | Improved |
| Add context frames to reconstruction rather than using none (ablation) | IV-CL · CATER | accuracy (%): 65.35% -> 70.14% | Improved |
| Use 8 rather than 4 context frames (ablation) | IV-CL · CATER | accuracy (%): 67.47% -> 70.14% | Improved |
| Use 8 rather than 16 context frames (ablation) | IV-CL · CATER | accuracy (%): 64.34% -> 70.14% | Improved |
Closest papers by meaning
- Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · similarity 0.88)
- Latent Implicit Visual Reasoning (CVPR 2026 · similarity 0.88)
- Train the Agent, Not the Expert: Learning to Harness Heterogeneous Experts for Multi-Turn Visual Reasoning (NeurIPS 2026 · similarity 0.87)
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.87)
- Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.87)
- Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization (CVPR 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.
