Add README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# DeltaVLM
|
| 2 |
+
|
| 3 |
+
This repository hosts the pretrained DeltaVLM checkpoint from the paper [DeltaVLM: Interactive Remote Sensing Image Change Analysis via Instruction-guided Difference Perception](https://arxiv.org/abs/2507.22346).
|
| 4 |
+
|
| 5 |
+
## Contents
|
| 6 |
+
|
| 7 |
+
- `checkpoint_best.pth`: pretrained DeltaVLM checkpoint
|
| 8 |
+
- `configs/evaluate.yaml`: example evaluation config used in the project codebase
|
| 9 |
+
|
| 10 |
+
## Important note
|
| 11 |
+
|
| 12 |
+
This checkpoint does **not** include the frozen base LLM weights. The original project loads a Vicuna-7B-v1.5-compatible model separately and then applies the DeltaVLM checkpoint on top.
|
| 13 |
+
|
| 14 |
+
In the local reproduction environment, the Vicuna path was substituted with a compatible `llavav1.5-7b` directory for loading.
|
| 15 |
+
|
| 16 |
+
## Code and dataset
|
| 17 |
+
|
| 18 |
+
- Code: https://github.com/hanlinwu/DeltaVLM
|
| 19 |
+
- Dataset annotations: https://huggingface.co/datasets/hlwu/changechat-105k
|
| 20 |
+
- Image source: https://github.com/Chen-Yang-Liu/LEVIR-CC-Dataset
|
| 21 |
+
|
| 22 |
+
## Example
|
| 23 |
+
|
| 24 |
+
After cloning the code repo and preparing the required base model plus dataset files, evaluation can be run with:
|
| 25 |
+
|
| 26 |
+
```bash
|
| 27 |
+
python infer_subtasks.py --cfg_path configs/evaluate.yaml --n_samples 2
|
| 28 |
+
```
|
| 29 |
+
|
| 30 |
+
## License
|
| 31 |
+
|
| 32 |
+
Please follow the license terms of this project repository, the referenced dataset, and the required upstream base model.
|