CSE5519 Advances in Computer Vision (Topic J: 2023 - 2024: Open-Vocabulary Object Detection)
Grounding DINO
Novelty in Grounding DINO
- Use CLIP to enhance the feature with DETER
- Contrastive loss for text-region alignment
- Localization loss-box regression (DINO style)
- Auxiliary loss across decoder layers
Top 900 bounding boxes for inference.
Tip
This paper shows a novel approach to open-vocabulary object detection by marrying DINO with CLIP. The authors use a DINO model to get the query features and then use a grounding head to get the bounding box and class label.
I’m really interested in the number of bounding boxes for inference. I wonder how fine-grained the bounding boxes are? Does it serve a good reference for counting problems and doing logical reasoning for example the hand with 6 fingers?
Last updated on