A naukri.com initiative
Towards Data Science
1M
156
Adding Training Noise To Improve Detections In Transformers
- Modern vision transformers are utilizing noise addition to enhance the performance of object detection tasks.
- Early vision transformers like DETR used learned decoder queries for object detection, but had slow convergence.
- Recent transformer architectures have implemented deformable aggregation and spatial anchors for improved detection results.
- The Hungarian algorithm is used for prediction to ground truth matching in transformers, leading to unstable training objectives.
- DN-DETR addresses the unstable matching issue by introducing noise to ground truth boxes, improving model stability and convergence speed.
- DINO enhances denoising by incorporating contrastive learning, improving detection performance even further.
- Temporal models like Sparse4Dv3 leverage denoising and temporal denoising groups for object tracking across frames.
- Denoising in vision transformers accelerates convergence and boosts detection results, especially with learnable anchors.
- The use of denoising raises questions about the necessity of learnable anchors and the impact on models with non-learnable anchors.
- While denoising improves stability in gradient descent, the relevance in models with spatially constrained queries remains a topic for further exploration.
Read Full Article
10 Likes
For uninterrupted reading, download the app