About cameltrack_train.yaml

[zhangzihao0916]

This is the content of the cameltrack_train.yaml file (configuration for training). In this file, the modules.tracker_states loads the .pklz file for the purpose of generating training data. What is the purpose of the state.load_file loading the .pklz file?

[VlSomers]

You are right for the modules.tracker_states loading the .pklz file for the purpose of generating training data (i.e. training data containing track/detection pairs to train the CAMEL module). The other .pklz in the state.load_file contains all the data required to perform multi-object tracking at inference (bboxes, ReID embeddings, keypoints, etc). It is a way to perform inference faster, so that only the association module (i.e. CAMEL) is run, without requiring to run the other modules (detection, reid, pose) again and again.

[zhangzihao0916]

You are right for the modules.tracker_states loading the .pklz file for the purpose of generating training data (i.e. training data containing track/detection pairs to train the CAMEL module). The other .pklz in the state.load_file contains all the data required to perform multi-object tracking at inference (bboxes, ReID embeddings, keypoints, etc). It is a way to perform inference faster, so that only the association module (i.e. CAMEL) is run, without requiring to run the other modules (detection, reid, pose) again and again.你说得对，modules.tracker_states加载.pklz文件是为了生成训练数据（即包含轨迹/检测对的训练数据，用于训练CAMEL模块）。state.load_file中的其他.pklz文件包含了在推理时执行多目标跟踪所需的所有数据（边界框、ReID嵌入、关键点等）。这是一种加快推理速度的方法，这样就只需要运行关联模块（即CAMEL），而无需反复运行其他模块（检测、重识别、姿态估计）。

Your explanation has been extremely helpful to me! I have one more request: as a beginner, training the model from scratch is quite challenging. Could you please enrich the documentation for training the model from scratch, starting from obtaining the CAMEL dataset?