STAR

[luckytanyy]

The pixel sizes of the data in the STAR dataset vary. Do we need to preprocess the data? If data preprocessing is required, can you provide the relevant code for data preprocessing.

[yangxue0827]

Basically, the data preprocessing is the same as DOTA.
Note that for STAR, the validation set does not need to be processed because we provide online large-scale image evaluation.

[mason1008]

@yangxue0827 “Note that for STAR, the validation set does not need to be processed because we provide online large-scale image evaluation.”

1. Training is consistent with DOTA, but what are the specific differences between validation and DOTA? How should this be implemented in detail?
2. When will the code for HOD-Net be released as open source?

Thank you

[yangxue0827]

1. Refer to test set in config.
   https://github.com/yangxue0827/STAR-MMRotate/blob/08a475bd0f6cf2ce919c99cc75093fafa5278d6a/configs/h2rbox_v2p/h2rbox_v2p_r50_fpn_1x_star_le90.py#L102-L109
2. It will be released after the paper accepted.

[mason1008]

1. I used tools/data/dota/split/img_split.py to process the STAR dataset during both training and testing. When training with oriented_rcnn_r50_fpn_1x_star_le90.py, there is no difference in the results between test=dict(type='STARWSOODDataset', ...) and test=dict(type='STARDataset', ...) when train=dict(type='STARDataset', ...) is used.
2. In Table 2 of the paper, the mAP of HOD-Net is 43.6 and 55.9, corresponding to single-scale or multi-scale for DOTA ?
3. In Table 3 of the paper, is the only difference between Resizing and Cropping the use of model = dict(type='OrientedRCNN', ...) and model = dict(type='OrientedRCNNCrop', ...) ?

[yangxue0827]

1. STARWSOODDataset is for weakly supervised detectors (such as H2RBox-v2), and STARDataset is for fully supervised detectors.

2. Multi-scale, involving data multi-scale cropping

3. OrientedRCNNCrop is specifically for STAR data. The difference from OrientedRCNN is that the former can support online evaluation of large images, so there is no need to crop the validation/test set.

[luckytanyy]

I would like to ask a question about the test set. If the test set is cropped 1024x1024 and uploaded to the evaluation website, and the result is empty, can the test set be cropped?