Research map / Detection and segmentation / Weakly supervised detection and segmentation
Instance Localization for Self-Supervised Detection Pretraining
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| pretrain with instance localization using pasted foregrounds and bounding-box region features instead of holistic instance discrimination | MoCo-v2, ResNet50-C4 · PASCAL VOC test2007 | AP: 57.4 -> 58.4 | Helped |
| pretrain with instance localization using pasted foregrounds and bounding-box region features instead of holistic instance discrimination | MoCo-v2, ResNet50-C4 · COCO val2017, Mask R-CNN 2× | APbb: 41.2 -> 41.8 | Helped |
| pretrain with instance localization using pasted foregrounds and bounding-box region features instead of holistic instance discrimination | MoCo-v2, ResNet50-FPN · COCO val2017, Mask R-CNN 2× | APbb: 42.5 -> 43.3 | Helped |
| pretrain with instance localization | MoCo-v2, ResNet50-C4 · Mini COCO val2017, 10% train | APbb: 22.7 -> 26.0 | Helped |
| add RoIAlign for architectural alignment (ablation) | MoCo-v2, ResNet50-FPN · COCO val2017, 1× ablation | APbb: 39.8 -> 40.2 | Helped |
| add copy-paste instance localization and bounding-box representations (ablation) | RoIAlign baseline, ResNet50-FPN · COCO val2017, 1× ablation | APbb: 40.2 -> 41.1 | Helped |
| add bounding-box augmentation using spatially jittered boxes (ablation) | Instance localization, ResNet50-FPN · COCO val2017, 1× ablation | APbb: 41.1 -> 41.4 | Helped |
| use instance localization rather than holistic instance discrimination | MoCo-v2, ResNet50-C4 · ImageNet linear probe | classification accuracy: 67.7 -> 61.7 | Hurt |
| use instance localization rather than holistic instance discrimination | MoCo-v2, ResNet50-C4 · Linear localization evaluation | localization accuracy: 71.9 -> 74.2 | Helped |
| increase InsLoc pretraining from 200 to 400 epochs | InsLoc, ResNet50-C4 · COCO val2017, Mask R-CNN 2× | APbb: 41.4 -> 41.8 | Helped |
| increase InsLoc pretraining from 200 to 400 epochs | InsLoc, ResNet50-FPN · COCO val2017, Mask R-CNN 2× | APbb: 43.2 -> 43.3 | No clear effect |
Closest papers by meaning
- Self-Supervised Learning of Intertwined Content and Positional Features for Object Detection (ICML 2025 · similarity 0.88)
- Training Object Class Detectors With Click Supervision (CVPR 2017 · similarity 0.88)
- Multi-Task Self-Supervised Object Detection via Recycling of Bounding Box Annotations (CVPR 2019 · similarity 0.87)
- Self-Supervised Object Localization with Joint Graph Partition (AAAI 2022 · similarity 0.87)
- Learning Instance Activation Maps for Weakly Supervised Instance Segmentation (CVPR 2019 · similarity 0.87)
- Unsupervised Object Detection Pretraining with Joint Object Priors Generation and Detector Learning (NeurIPS 2022 · 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.