Research map / Detection and segmentation / Weakly supervised detection and segmentation
Box-based Refinement for Weakly Supervised and Unsupervised Localization Tasks
Eyal Gomel, Tal Shaharbany, Lior Wolf
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 box-union loss with probability 0.5 when refining g (ablation) | refined phrase grounding network · Flickr30K, MS-COCO trained, bbox accuracy | bbox accuracy (%): 28.74 -> 47.26 | Improved |
| add the original g training losses alongside the detector loss (ablation) | refined phrase grounding network · Flickr30K, MS-COCO trained, pointing accuracy | pointing accuracy (%): 68.47 -> 77.19 | Improved |
| predict k=10 boxes instead of k=1 in phrase grounding (ablation) | refined phrase grounding network · Flickr30K, MS-COCO trained, bbox accuracy | bbox accuracy (%): 32.35 -> 47.26 | Improved |
| refine g with detector-based loss | WWbL phrase grounding · Flickr30K, MS-COCO trained, bbox accuracy | bbox accuracy (%): 35.75 -> 47.26 | Improved |
| add regularization terms to detector loss in object discovery (ablation) | refined DINO with LOST · VOC07 CorLoc | CorLoc (%): 61.72 -> 62.0 | Improved |
| use k=1 box instead of k=10 in object discovery (ablation) | refined DINO with LOST · VOC07 CorLoc | CorLoc (%): 62.0 -> 62.54 | Improved |
| vary k from 1 to 15 in object discovery (ablation) | refined DINO with LOST · MS-COCO20K CorLoc | CorLoc (%): 52.00 (k=1), 51.98 (k=10), 50.60 (k=15) | No clear effect |
| refine DINO with detector loss on MOVE boxes | MOVE · VOC07 CorLoc | CorLoc (%): 76.0 -> 77.5 | Improved |
| refine DINO with detector loss on TokenCut boxes | TokenCut · VOC07 CorLoc | CorLoc (%): 68.8 -> 69.0 | No clear effect |
| refine DINO with detector loss on LOST boxes | LOST · VOC07 CorLoc | CorLoc (%): 61.9 -> 62.0 | No clear effect |
| add refinement on top of MOVE + CAD | MOVE + CAD · VOC07 CorLoc | CorLoc (%): 77.1 -> 78.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MOVE — MOVE: Unsupervised Movable Object Segmentation and Detection (NeurIPS 2022) · better on 3 of 3 shared results
- TokenCut — TokenCut: Segmenting Objects in Images and Videos With Self-Supervised Transformer and Normalized Cut (TPAMI 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Training Object Class Detectors With Click Supervision (CVPR 2017 · similarity 0.89)
- Pairwise Similarity Knowledge Transfer for Weakly Supervised Object Localization (ECCV 2020 · similarity 0.88)
- Weakly Supervised Object Localization with Multi-Fold Multiple Instance Learning (TPAMI 2016 · similarity 0.88)
- Weakly Supervised Object Localization Using Size Estimates (ECCV 2016 · similarity 0.88)
- Learning to Localize Objects with Noisy Labeled Instances (AAAI 2019 · similarity 0.88)
- Rethinking the Route Towards Weakly Supervised Object Localization (CVPR 2020 · similarity 0.88)
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.
