Research map / Label-efficient and robust learning / Recognition and long-tail
Evidential Active Recognition: Intelligent and Prudent Open-World Embodied Perception
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Evidential Active Recognition, which models active object recognition as sequential evidence gathering with Dirichlet-based evidential deep learning, Dempster-Shafer fusion across views, and an uncertainty-aware belief reward for PPO policy training, evaluated on a new MP3D-based dataset of 13,200 instances.
Problem
Active recognition agents are built under the closed-world assumption, so their softmax-based recognition and rewards are misleading when the target is absent or ambiguous in open-world embodied settings.
Approach
Replace softmax outputs with evidential deep learning (Dirichlet belief and uncertainty per view), fuse per-step opinions late via Subjective Logic/Dempster-Shafer combination, and use the belief for the correct class as the policy reward so that uncertain or target-absent views are not rewarded.
Key results
- Ours achieves 64.4% top-1 and 84.3% top-3 overall success at t=10, versus 57.2% top-1 for Amodal-Rec and 61.8% top-1 for Ours with Fixation.
- On hard instances the method improves top-1 by +8.3 points from t=1 to t=10, compared with +2.0 on easy and +2.6 on moderate.
- Removing the uncertainty-aware reward reduces overall top-1 from 64.4% to 61.6%, with the largest drop on hard instances (58.0% to 53.4%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use active movement with uncertainty-aware policy learning and evidence fusion instead of the Fixation policy | Ours · Easy | top-1 success rate: 66.2 -> 69.9 | Helped |
| use active movement with uncertainty-aware policy learning and evidence fusion instead of the Fixation policy | Ours · Moderate | top-1 success rate: 58.0 -> 59.7 | Helped |
| use active movement with uncertainty-aware policy learning and evidence fusion instead of the Fixation policy | Ours · Hard | top-1 success rate: 56.1 -> 58.0 | Helped |
| use active movement with uncertainty-aware policy learning and evidence fusion instead of the Fixation policy | Ours · All | top-1 success rate: 61.8 -> 64.4 | Helped |
| add evidence fusion rather than Max-prediction (ablation) | Ours · All | top-1 success rate: 62.9 -> 64.4 | Helped |
| add evidence fusion rather than Last-step prediction (ablation) | Ours · All | top-1 success rate: 60.7 -> 64.4 | Helped |
| add evidence fusion rather than Average (ablation) | Ours · All | top-1 success rate: 63.0 -> 64.4 | Helped |
| add evidence fusion rather than Vote (ablation) | Ours · All | top-1 success rate: 62.8 -> 64.4 | Helped |
| add uncertainty-aware reward during policy learning (ablation) | Ours · All | top-1 success rate: 61.6 -> 64.4 | Helped |
| add uncertainty-aware reward during policy learning (ablation) | Ours · Hard | top-1 success rate: 53.4 -> 58.0 | Helped |
Surprising findings
- Amodal-Rec with a trained policy performs roughly the same as its Single-View baseline, suggesting feature-level temporal aggregation struggles in open-world active recognition.
- Random movement drops top-1 success by 7.5 points relative to Single-View for the proposed agent, showing that uninformed movement can hurt.
Limitations
- Fixed number of steps (T=10) rather than learned stopping, and the evaluation metric does not account for number of observations.
- Ground-truth bounding box queries are used for the target mask instead of a tracker, so the tracking uncertainty is not evaluated.
- Evaluation is limited to 27 indoor object categories in simulation.
Experimental setup
Datasets: HM3D MP3D
Benchmarks: MP3D-based active recognition test set
Metrics: top-1 success rate top-3 success rate change in success rate from t=1 to t=10
Baselines: Single-View Random Fixation Amodal-Rec
Builds on: evidential deep learning Subjective Logic PPO Amodal-Rec Faster R-CNN
Closest papers by meaning
- PhysVLM-AVR: Active Visual Reasoning for Multimodal Large Language Models in Physical Environments (NeurIPS 2025 · similarity 0.85)
- Active Open-Vocabulary Recognition: Let Intelligent Moving Mitigate CLIP Limitations (CVPR 2024 · similarity 0.85)
- I-Perceive: A Foundation Model for Vision-Language Active Perception (NeurIPS 2026 · similarity 0.85)
- ACTIVE-o3 : Empowering MLLMs with Active Perception via Pure Reinforcement Learning (ICML 2026 · similarity 0.85)
- Mind the GAP: Glimpse-based Active Perception improves generalization and sample efficiency of visual reasoning (ICLR 2025 · similarity 0.84)
- Active Test-time Vision-Language Navigation (NeurIPS 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.