Research map / Detection and segmentation / Weakly supervised detection and segmentation
SODA10M: A Large-Scale 2D Self/Semi-Supervised Object Detection Dataset for Autonomous Driving
Jianhua Han, Xiwen Liang, Hang Xu, Kai Chen, Lanqing HONG, Jiageng Mao, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Xiaodan Liang, Chunjing Xu
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 |
|---|---|---|---|
| pretrain MoCo-v1 on SODA10M instead of BDD100K | MoCo-v1 · SODA10M val (OD) | mAP: 37.1 -> 38.9 | Improved |
| pretrain on BDD100K (smaller, more diverse) instead of nuScenes | MoCo-v1 · SODA10M val (OD) | mAP: 36.2 -> 37.1 | Improved |
| use dense contrastive loss (DenseCL) instead of MoCo-v1 when pretraining on SODA10M | Faster RCNN fine-tune · SODA10M val | mAP: 38.9 -> 38.1 | Hurt |
| use dense contrastive loss (DenseCL) instead of MoCo-v1 when pretraining on ImageNet | Faster RCNN fine-tune · SODA10M val | mAP: 39.0 -> 39.9 | Improved |
| pretrain on SODA10M instead of ImageNet | MoCo-v1 · SODA10M val, Faster-RCNN | mAP: 39.0 -> 38.9 | No clear effect |
| add 5M extra pre-training images (ablation) | SimCLR on SODA10M · SODA10M val, Faster-RCNN | mAP: 35.9 -> 37.1 | Improved |
| add 5M extra pre-training images (ablation) | MoCo-v2 on SODA10M · SODA10M val, Faster-RCNN | mAP: 38.7 -> 38.6 | No clear effect |
| increase pseudo-labeling unlabeled data from 50K to 100K (ablation) | Pseudo Labeling · SODA10M val | mAP: 39.3 -> 39.9 | Improved |
| increase pseudo-labeling unlabeled data from 100K to 500K (ablation) | Pseudo Labeling · SODA10M val | mAP: 39.9 -> 38.5 | Hurt |
| add consistency regularization (STAC) to pseudo labeling | Faster RCNN semi-supervised · SODA10M val | mAP: 39.9 -> 42.8 | Improved |
| use Unbiased Teacher (EMA + focal loss) instead of STAC | Faster RCNN semi-supervised · SODA10M val | mAP: 42.8 -> 46.2 | Improved |
| use semi-supervised Unbiased Teacher instead of supervised only | Faster RCNN · SODA10M night domain | mAP: 21.1 -> 39.7 | Improved |
Closest papers by meaning
- One Million Scenes for Autonomous Driving: ONCE Dataset (NeurIPS 2021 · similarity 0.90)
- Scalability in Perception for Autonomous Driving: Waymo Open Dataset (CVPR 2020 · similarity 0.88)
- CODA: A Real-World Road Corner Case Dataset for Object Detection in Autonomous Driving (ECCV 2022 · similarity 0.88)
- SelfD: Self-Learning Large-Scale Driving Policies From the Web (CVPR 2022 · similarity 0.88)
- MultiSiam: Self-Supervised Multi-Instance Siamese Representation Learning for Autonomous Driving (ICCV 2021 · similarity 0.88)
- AIDE: An Automatic Data Engine for Object Detection in Autonomous Driving (CVPR 2024 · 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.
