Train on a Custom Dataset

megvii-research / MOTR

[ECCV2022] MOTR: End-to-End Multiple-Object Tracking with TRansformer

Other

613 stars 93 forks source link

Train on a Custom Dataset #67

Open abdallah1989203 opened 1 year ago

abdallah1989203 commented 1 year ago

My dataset looks like this: ..data/vid1/images, data/vid1/labels, vid1...vidx Each vid represents a video. How can I convert my dataset to MOT format? CVAT offers the MOT 1.1 export format, but it is missing the seqinfo.ini file for each vid. If I have multiple vids, how can I train a model on them?

abdallah1989203 commented 1 year ago

@zyayoung, @dbofseuofhust can you help please?

subbubhat99 commented 2 months ago

Hi abdallah1989203! Could you let me know if you figured out a way to create the MOT dataset? I have a bunch of videos, only some of them being annotated through all frames, and the rest yet to be labeled. I've looked into Google cloud's Vertex AI to create the object tracking dataset but it seems too non-trivial for starters! Kindly let me know how you fixed it! Regards, Subramanya

abdallah1989203 commented 2 months ago

Hi @subbubhat99

As far as I remember, I created the seq.info file manually, i.e., automatically through a script, so that there is a seq.info file for each video.