search

nwojke / cosine_metric_learning

Deep Cosine Metric Learning for Person Re-identification
GNU General Public License v3.0
586 stars 201 forks source link

Does the mars-small128.pb file provided by the author only use the network in cosine_metric_learning_master for training? #35

Open ZhaoMonica opened 5 years ago

ZhaoMonica commented 5 years ago

Hello. Does the mars-small128.pb file you provided only use the network in cosine_metric_learning_master for training? Have you made any changes? I saw you say it's better to look for network files than in deep sort, but I didn't find the corresponding network files in deep sort.

nwojke commented 5 years ago

The checkpoint file that ships with the deep sort repository has some minor differences compared to the version in this repository. The deep sort repository uses a frozen inference graph for feature extraction (the .pb file), allowing the use of both, the old or the newer network version. The deep sort network file is here (also linked in the README of Deep SORT).

Let me know if you have other questions.

cydni commented 2 years ago

@nwojke Quick question about training a custom feature extractor for a soccer ball data set. My goal is to improve my ball tracker performance for soccer broadcast videos as it is currently using the mars.pb model. For MAX_LABEL, should this value be 1 (since the only class I am tracking is ball) or the number of balls in the data set (15, since that's the number of games I annotates). I am not sure if this will make a difference since there will typically only be one ball on screen to track at a time and they all look very similar.