Datasets:
Dataset preparation issues: Missing category mapping files and broken Obj365 image source
Thanks for releasing WildDet3D-Data. I noticed two dataset preparation issues:
The README lists
annotations/InTheWild_v3_*_class_map.jsonas expected category mapping files, but these files do not appear to be included in the dataset download. Could you confirm whether they are required for training, and if so, upload them or document how to generate them from the annotation JSONs?The README instructs users to download Objects365 images from the official Objects365 website, but the official download links appear to be unavailable now. Is there an alternative source, mirror, or recommended procedure for obtaining the Obj365 training images used by WildDet3D-Data?
These two points make it difficult to fully reproduce the expected dataset layout from the current instructions.
Hi,
Sry for the confusion, the class_map will be automatically generated through the training, and you can also easily extract it from the annotation JSON files.
And for Object365, we are using the v1 version of it. We are not allowed to directly share this due to the license, but you might find it on some third-party repos like https://huggingface.co/datasets/jxu124/objects365/tree/main/data (non-verified)
Got it. Thank you for your reply!