Research map / Reinforcement learning / Visual and representation RL
AdaGlimpse: Active Visual Exploration with Arbitrary Glimpse Position and Scale
Adam Pardyl, Michał Wronka, Maciej Wołczyk, Kamil Adamczewski, Tomasz Trzcinski, Bartosz Zieliński
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 |
|---|---|---|---|
| select glimpses at arbitrary position and scale instead of fixed-scale grids | AdaGlimpse · ImageNet-1k classification | accuracy (%): 76.13 -> 77.54 | Improved |
| add early stopping based on predicted class probability | AdaGlimpse · ImageNet-1k classification | pixel percentage (%): 28.57 -> ~16.94 | Improved |
| use adaptive scale glimpses instead of simple grids at same pixel budget | AdaGlimpse · ImageNet-1k reconstruction (6% pixels) | RMSE: 20.9 vs 23.4 (AME, 18.75% pixels, SUN360 best baseline) | Improved |
| adaptive glimpses versus AME at similar pixel budget | AdaGlimpse · ADE20K segmentation | PA (%): 70.3 -> 70.0 | No clear effect |
| replace glimpse coordinates in RL state with dataset mean (ablation) | AdaGlimpse RL agent · ImageNet-1k classification | accuracy (%): 68.25 -> 77.54 | Improved |
| include transformer latent in RL state (ablation) | AdaGlimpse RL agent · ImageNet-1k classification | accuracy (%): 61.82 -> 77.54 | Improved |
| include patch pixels in RL state (ablation) | AdaGlimpse RL agent · ImageNet-1k classification | accuracy (%): 76.99 -> 77.54 | Improved |
| include patch importance in RL state (ablation) | AdaGlimpse RL agent · ImageNet-1k classification | accuracy (%): 77.36 -> 77.54 | Improved |
| train only on ImageNet-1k and test zero-shot on other reconstruction datasets | AdaGlimpse · SUN360/ADE20K/COCO reconstruction | RMSE: 11.1 vs 23.6 (SUN360, AME) | Improved |
Closest papers by meaning
- Glimpse-Attend-and-Explore: Self-Attention for Active Visual Exploration (ICCV 2021 · similarity 0.91)
- Sidekick Policy Learning for Active Visual Exploration (ECCV 2018 · similarity 0.89)
- Symmetry-Aware Neural Architecture for Embodied Visual Exploration (CVPR 2022 · similarity 0.87)
- An Interactive Navigation Method with Effect-oriented Affordance (CVPR 2024 · similarity 0.85)
- ACTIVE-o3 : Empowering MLLMs with Active Perception via Pure Reinforcement Learning (ICML 2026 · similarity 0.85)
- Leveraging Separated World Model for Exploration in Visually Distracted Environments (NeurIPS 2024 · 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.
