search

ifzhang / FairMOT

[IJCV-2021] FairMOT: On the Fairness of Detection and Re-Identification in Multi-Object Tracking
MIT License
4.03k stars 934 forks source link

Partial Human Body Tracking #544

Open hust-zXX opened 1 year ago

hust-zXX commented 1 year ago

When I applied it, I found that if the person squatted down or only exposed the upper body, the target could not be effectively tracked. Is there any quick way to improve it?

janthmueller commented 1 year ago

As far as the detection part is concerned, the cause is probably a too small number of training examples that have these features. I believe that, for example, static individuals are more likely to have these features, but are filtered out within datasets such as MOT16 and MOT20: image MOT16: A Benchmark for Multi-Object Tracking This also affects the re-id part, as it may reduce the similarity between, for example, a person standing and squatting. To improve this part, the appropriate way would be to finetune the model. Maybe an improvement could be achieved if static persons are not filtered out of the datasets.

Another option would be to use better alternatives to the kalman filter and hungarian algorithm.

hust-zXX commented 1 year ago

As far as the detection part is concerned, the cause is probably a too small number of training examples that have these features. I believe that, for example, static individuals are more likely to have these features, but are filtered out within datasets such as MOT16 and MOT20: image MOT16: A Benchmark for Multi-Object Tracking This also affects the re-id part, as it may reduce the similarity between, for example, a person standing and squatting. To improve this part, the appropriate way would be to finetune the model. Maybe an improvement could be achieved if static persons are not filtered out of the datasets.

Another option would be to use better alternatives to the kalman filter and hungarian algorithm.

thanks,I will try it