search

rohanchandra30 / TrackNPred

This is the code base for our ACM CSCS 2019 paper: "RobustTP: End-to-End Trajectory Prediction for Heterogeneous Road-Agents in Dense Traffic with Noisy Sensor Inputs". This codebase contains implementations for several trajectory prediction methods including Social-GAN and TraPHic.
https://arxiv.org/pdf/1907.08752.pdf
196 stars 63 forks source link

Question regardin DeepSort implementation #12

Open drapado opened 4 years ago

drapado commented 4 years ago

Hi, I have one question regarding the features for the deep association metric on DeepSort. I've noticed you used the CNN to get the features used in DeepSort wich was trained in the MARS dataset. This dataset is meant for person re-identification. Then my question is: How do you deal with object of different classes? I mean, do you use the same network (trained for person re-identification) to generate the deep feature descriptor of objects such as cars?

Thanks in advance!

rohanchandra30 commented 4 years ago

We used the same network as we observed fairly decent tracking results without re-training the network. Maybe 1-2 % improvement occurs by re-training, but we didn't require it.

drapado commented 4 years ago

Hi @rohanchandra30, Thanks for your reply, interesting results. I'm trying to use deepsort with COCO objects, but my experiments show that the original feature extraction network trained con MARS is not able to individually differentiate between different objects. From my experience, it seems to rely a lot on the color of the object, and for a lot of objects that's not useful.