Research map / Label-efficient and robust learning / Self-supervised and contrastive
Transitive Invariance for Self-Supervised Visual Representation Learning
Xiaolong Wang, Kaiming He, Abhinav Gupta
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 |
|---|---|---|---|
| use transitive graph with both tracking and context invariance instead of tracking only | Triplet-Siamese network · VOC 2007 test (Fast R-CNN) | mAP: 60.2 -> 63.2 | Improved |
| use transitive graph instead of context prediction only | Triplet-Siamese network · VOC 2007 test (Fast R-CNN) | mAP: 61.5 -> 63.2 | Improved |
| use transitive graph instead of multi-task learning of both tasks | Triplet-Siamese network · VOC 2007 test (Fast R-CNN) | mAP: 62.1 -> 63.2 | Improved |
| use 30-frame tracks instead of 15-frame tracks (ablation) | Triplet-Siamese network · VOC 2007 test (Fast R-CNN) | mAP: 61.5 -> 63.2 | Improved |
| cluster with learned still-image features instead of HOG (ablation) | Triplet-Siamese network · VOC 2007 test (Fast R-CNN) | mAP: 60.4 -> 63.2 | Improved |
| add self-supervised pre-training versus training from scratch | Fast R-CNN · VOC 2007 test (Fast R-CNN) | mAP: 39.7 -> 63.2 | Improved |
| use transitive self-supervised pre-training instead of ImageNet | Fast R-CNN · VOC 2007 test (Fast R-CNN) | mAP: 67.3 -> 63.2 | Hurt |
| use transitive self-supervised pre-training instead of context pre-training | Faster R-CNN · VOC 2007 test (Faster R-CNN, all layers fine-tuned) | mAP: 62.6 -> 65.0 | Improved |
| freeze all conv layers during fine-tuning with ours vs context | Faster R-CNN · VOC 2007 test (Faster R-CNN, conv1-conv5 frozen) | mAP: 38.1 -> 43.1 | Improved |
| use multi-task alternative | Faster R-CNN · VOC 2007 test (Faster R-CNN, conv1-conv5 frozen) | mAP: 43.1 -> 27.6 | Hurt |
| use self-supervised pre-training instead of ImageNet | Faster R-CNN · COCO minival | AP: 24.4 -> 23.5 | Hurt |
| use self-supervised pre-training instead of shorter-schedule ImageNet | Faster R-CNN · COCO minival | AP: 23.7 -> 23.5 | No clear effect |
Closest papers by meaning
- Demystifying Contrastive Self-Supervised Learning: Invariances, Augmentations and Dataset Biases (NeurIPS 2020 · similarity 0.91)
- Self-Supervised Learning of Video-Induced Visual Invariances (CVPR 2020 · similarity 0.90)
- Towards Efficient and Effective Self-Supervised Learning of Visual Representations (ECCV 2022 · similarity 0.90)
- Exploring the Equivalence of Siamese Self-Supervised Learning via a Unified Gradient Framework (CVPR 2022 · similarity 0.89)
- In-Context Symmetries: Self-Supervised Learning through Contextual World Models (NeurIPS 2024 · similarity 0.88)
- Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency (CVPR 2025 · 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.
