What is the difference between coco_2014_train_nonvoc_with_small train dataset in single_branch_pretraining_coco_pvt_v2_b2_li.yaml and coco_2014_train_nonvoc train dataset in two_branch_training_coco_pvt_v2_b2_li.yaml? Unfortunately, I did not find how to generate the coco_2014_train_nonvoc_with_small train dataset.

[Alan-lab]

[GuangxingHan]

Thanks for pointing out this issue, and we have uploaded the python file 1_split_filter_with_small.py.

The with_small suffix denotes the dataset include the small scale objects (bbox_area < 32 * 32). We find that small objects are useful to train the single branch models with FPN, but they does not work well with our two-branch models. How to improve small object detection for two-branch models is an interesting future work.

[Alan-lab]

Oops! This is indeed a noteworthy point.