Research map / Detection and segmentation / Open-set and few-shot detection
Reward Finetuning for Faster and More Accurate Unsupervised Object Discovery
Katie Z Luo, Zhenzhen Liu, Xiangyu Chen, Yurong You, Sagie Benaim, Cheng Perng Phoo, Mark Campbell, Wen Sun, Bharath Hariharan, Kilian Q Weinberger
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 repeated self-training with reward-ranked local exploration and finetuning | PointRCNN object discovery · Lyft test, 0-80m | mAP IoU 0.5: 39.6 -> 45.3 | Improved |
| replace repeated self-training with reward-ranked local exploration and finetuning | PointRCNN object discovery · Lyft test, 0-80m | mAP IoU 0.7: 18.8 -> 29.6 | Improved |
| replace repeated self-training with reward-ranked finetuning | PointRCNN object discovery · Ithaca365 test, 0-80m | mAP: 27.1 -> 35.1 | Improved |
| add exploration by locally sampling boxes around predictions (ablation) | DRIFT · Lyft test, 0-80m | mAP IoU 0.5: 0.0 -> 41.8 | Improved |
| add the shape prior reward (ablation) | DRIFT with filtering and alignment rewards · Lyft test, 0-80m | mAP IoU 0.5: 22.4 -> 38.3 | Improved |
| add alignment reward to filtering and shape prior (ablation) | DRIFT · Lyft test, 0-80m | mAP IoU 0.5: 0.0 -> 38.3 | Improved |
| increase exploration sample count from 50 to 200 at noise scale 0.3 (ablation) | DRIFT · Lyft test, 0-80m | mAP IoU 0.5: 32.9 -> 38.3 | Improved |
| increase exploration noise scale from 0.3 to 0.6 with 200 samples (ablation) | DRIFT · Lyft test, 0-80m | mAP IoU 0.5: 38.3 -> 0.6 | Hurt |
| add sampling near dynamic points to prediction-based exploration (ablation) | DRIFT · Lyft test, 0-80m | mAP IoU 0.5: 41.8 -> 39.3 | Hurt |
| add sampling from initial seed labels to exploration (ablation) | DRIFT · Lyft test, 0-80m | mAP IoU 0.5: 41.8 -> 39.5 | Hurt |
| change alignment reward μ_scale from 0.8 to 0.9 (ablation) | DRIFT · Lyft test, 0-80m | mAP IoU 0.5: 38.3 -> 38.1 | No clear effect |
| change alignment reward σ_scale from 0.2 to 0.1 (ablation) | DRIFT · Lyft test, 0-80m | mAP IoU 0.5: 38.3 -> 34.2 | Hurt |
Closest papers by meaning
- Reward Machines for Deep RL in Noisy and Uncertain Environments (NeurIPS 2024 · similarity 0.85)
- RLGF: Reinforcement Learning with Geometric Feedback for Autonomous Driving Video Generation (NeurIPS 2025 · similarity 0.85)
- Cycle Consistency Driven Object Discovery (ICLR 2024 · similarity 0.84)
- Leveraging Sub-Optimal Data for Human-in-the-Loop Reinforcement Learning (ICLR 2025 · similarity 0.84)
- Code as Reward: Empowering Reinforcement Learning with VLMs (ICML 2024 · similarity 0.84)
- Efficient Exploration and Discriminative World Model Learning with an Object-Centric Abstraction (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.
