search

KostadinovShalon / UAVDetectionTrackingBenchmark

78 stars 17 forks source link

missing dataset configs #1

Closed fcakyon closed 3 years ago

fcakyon commented 3 years ago

Hello, thanks a lot for this comprehensive benchmark repo. Im a PhD student and want to use this repo to pursue my research.

In this training config, it seems that your are using the default coco_detection.py dataset config file. It uses CocoDataset as the dataset type. However that dataset type does not include the drone class. Have you written your custom dataset class for mmdetection or have you added that class to original coco dataset file?

Secondly, what is the input size you have been using in your experiments?

Would you mind sharing your dataset config files so that it is easier for us to replicate the results?

KostadinovShalon commented 3 years ago

Hi,

The CocoDataset class is a subclass of CustomDataset. As you can see here, if you define a set of classes, it will override the default list of classes (although we used the models trained on the COCO Dataset as initialization). So no, in this case, we did not write a custom dataset class.

Regarding the input size for Faster RCNN, we used the default (1333, 800) as you point out (same with DETR, SSD: 512 x 512 and YOLOv3: 608 x 608).

fcakyon commented 3 years ago

Thanks a lot for clearing things up.