xView training and result reproduction

[nalemadi]

Hello @fcakyon,
Congratulations on your achievement and thank you for sharing the great job.

I was able to reproduce the results announced with Visdrone dataset by converting the dataset to coco format, training, and inference, which is excellent.
But I have faced some issues with training xView with the TOOD detector.

These are the steps I followed to train xView (similar to the actions performed on training Visdrone):

1. I converted xView dataset to coco using this script https://github.com/fcakyon/small-object-detection-benchmark/blob/main/xview/xview_to_coco.py which produces train.json (634 images) and val.json (212 images).
2. I used https://github.com/fcakyon/small-object-detection-benchmark/blob/main/xview/slice_xview.py to slice the val split images, which makes: val_300_0.json, val_300_025.json, val_400_0.json, val_400_025.json, val_500_0.json, val_500_025.json with its corresponding images directories.
3. I used this script to train the model using xView https://github.com/fcakyon/small-object-detection-benchmark/blob/main/mmdet_tools/train.py with this config file https://github.com/fcakyon/small-object-detection-benchmark/blob/main/mmdet_configs/xview_tood/tood_crop_300_500_cls_60.py (I only changed the BATCH_MULTIPLIER to lower value to fit into the available resources.
   I had the following error:

2023-03-28 22:58:06,076 - mmdet - ERROR - The testing results of the whole dataset is empty.

(Notes: I double-checked the converted val.json by performing the inference of xView using your checkpoint on TOOD+SF+SAHI+FI+PO setup and I was able to get the announced result. Also, I was able to train xView with https://github.com/fcakyon/small-object-detection-benchmark/blob/main/mmdet_configs/xview_tood/tood_full_cls_60.py config that got low results when using inference with TOOD+SF+SAHI+FI+PO setup)

I'm trying to figure out what went wrong. Could you please help me with that?
Looking forward to your reply

[fcakyon]

Hello @nalemadi

1. You should use the original training split during training (not the sliced dataset).
2. You can use the sliced val split during training.
3. You should use the eval script on the original dataset (not on the sliced dataset).

Is this what you are doing?

[nalemadi]

Yes, this is what I'm doing.