eval.md

Evaluation

1. FID

cd CC-Diff
cd eval

Install the required package:

pip install torch-fidelity

Then run:

bash fid.sh

2. YOLOScore

To evaluate YOLO Score, first convert the annotation format using dota2yolo.py in eval/utils:

python eval/utils/dota2yolo.py

After that, train the model according to the training process under the yoloscore directory.

Finally, replace the validation set (val) with your own generated validation results, and then run the evaluation.

3. Trainability

For the third metric, you need to first create the environment. You can refer to the environment setup under mmdet/docker.

Use covert_dota_coco_format.py to convert the annotations into COCO format.

For experiments with additional synthetic data, use combine_syn_images_dota.py to combine the data.

After the data preparation is complete, use the official tools provided by MMDetection (tools) for training and evaluation.