How gao-kun-lab/efficient-grounding-dino works
This repository presents Efficient Grounding DINO for remote-sensing visual grounding: a user supplies imagery and text, and the model grounds the text in the image with object predictions. The README documents RSVG/RSVG_HR training, augmentation, and fine-tuning; the sampled detector code shows prompt/token preparation, language-model encoding, and prediction. The model is also built on MMDetection. The grounding configs and dataset implementation are not present in the supplied tree, so their wiring is not mapped; the fusion and label-assignment details are documented at a high level rather than verified from source.
Main entry points: grounding_dino_Lquery.py Grounding DINO · tree_datasets.py Images and annotations · det_inferencer.py Detection inferencer · configs Training experiments
Components 8
Remote-sensing data
- Images and annotations
detection_data/tree_datasets.py - Text prompt
Grounding model
- Grounding DINO
mmdetection/mmdet/models/detectors/grounding_dino_Lquery.py - Language model
Inference runtime
- Detection inferencer
mmdetection/mmdet/apis/det_inferencer.py - Grounded predictions
- Training experiments
mmdetection/configs
Other components
- Researcher
Connections 8
- Researcher → Detection inferencer: submits inputs
- Detection inferencer → Grounding DINO: runs inference
- Grounding DINO → Language model: encodes prompts
- Text prompt → Grounding DINO: provides text
- Images and annotations → Grounding DINO: provides imagery
- Grounding DINO → Grounded predictions: returns grounding
- Images and annotations → Training experiments: training data
- Training experiments → Grounding DINO: configures experiments