Research map / Detection and segmentation / Open-set and few-shot detection
Spatially Consistent Representation Learning
Byungseok Roh, Wuhyun Shin, Ildoo Kim, Sungwoong Kim
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 |
|---|---|---|---|
| Use spatially matched local representations instead of global image-level matching | Self-supervised pretraining transferred to object detection · COCO val2017, Faster R-CNN and RetinaNet | AP: 40.0 -> 40.9 (Faster R-CNN); 37.7 -> 39.0 (RetinaNet) | Improved |
| Use spatially matched local representations instead of global image-level matching | Self-supervised pretraining transferred to localization tasks · PASCAL VOC test2007 detection | AP: 55.0 -> 57.2 | Improved |
| Increase the number of matched RoIs from 1 to 10 (ablation) | SCRL · COCO val2017 detection | AP: 39.2 -> 40.2 | Improved |
| Add exact spatial matching of the boxes rather than jittering them (ablation) | SCRL with 1 or 10 RoIs · COCO val2017 detection | AP: 39.2 -> 39.1 (10% jitter); 39.2 -> 38.8 (20% jitter); 40.2 -> 40.0 (10% jitter); 40.2 -> 39.6 (20% jitter) | Improved |
| Add IoU thresholding to reject redundant sampled boxes (ablation) | SCRL with 10 RoIs · COCO val2017 detection | AP: 39.8 -> 40.2 | Improved |
| Set pretraining learning rate to 0.3 instead of 0.45 (ablation) | SCRL · COCO val2017 detection | AP: 40.9 -> 41.2; +0.3 AP | Improved |
| Add negative pairs using the SimCLR framework (ablation) | SCRL upstream training · COCO detection | AP: +0.47 AP | Improved |
| Use input-level spatial consistency instead of feature-level matching (ablation) | SCRL-style pretraining · Localization downstream tasks | AP | Hurt |
| Pretrain with SCRL rather than BYOL | ResNet-50 · COCO detection with 10% training data | AP: 25.5 -> 26.4 | Improved |
| Pretrain with SCRL rather than BYOL | ResNet-50 · COCO val2017 detection schedules from 0.5× to 7× | AP: 37.0 -> 38.3; 40.0 -> 40.9; 41.6 -> 42.3; 41.7 -> 42.5; 41.6 -> 42.6 | Improved |
Closest papers by meaning
- RegionCL: Exploring Contrastive Region Pairs for Self-Supervised Representation Learning (ECCV 2022 · similarity 0.88)
- Region Similarity Representation Learning (ICCV 2021 · similarity 0.88)
- CSP: Self-Supervised Contrastive Spatial Pre-Training for Geospatial-Visual Representations (ICML 2023 · similarity 0.88)
- Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations (NeurIPS 2021 · similarity 0.87)
- Self-Supervised Image Representation Learning With Geometric Set Consistency (CVPR 2022 · similarity 0.87)
- Contrastive Attention Maps for Self-Supervised Co-Localization (ICCV 2021 · similarity 0.86)
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.
