YTEP-ZHI
commented
8 months ago @181860153 It will take a lot of engineering effort to adapt the codebase to other datasets, as it's heavily coupled with nuscenes annotations for multiple tasks throughout the data process, training, inference/validation. I suggest implementing a simpler version of uniad for your custom datasets. For example, remove these highly optimized submodules but keep the overall architectures with a bunch of attention as well as queries for interactions across different tasks.
Thank you for your awesome work! I would like to take some pictures to test the model. Besides six images taken from cam_front_left, cam_front, cam_front_right, cam_back_right, cam_back, cam_back_left, what else information should I give the model? And how should I arrange my data? Thank you very much!