search

PeizeSun / TransTrack

Multiple Object Tracking with Transformer
MIT License
627 stars 109 forks source link

About training #14

Closed lichunxiang93 closed 2 years ago

lichunxiang93 commented 3 years ago

thank you for your sharing，I have a few questions，Hope you can answer。 1、 I noticed that you have given the results of training using only the crowdHuman data set. I would like to ask whether this algorithm can achieve the purpose of tracking only by using the pedestrian detection data set training? 2、Can I use a large number of target detection data sets combined with a small number of tracking data sets for training？

PeizeSun commented 3 years ago

Hi~

  1. Of course. Please see "Evaluate pre-trained CrowdHuman model on MOT" in readme.
  2. In fact, our code is only based on only detection data to train a tracking model. The key is (1) random select an image (2) its adjacent frame image is simulated by randomly scaling and translating this image.