Train Re-ID network for multiple class on custom data

I want to train re-identification model on custom dataset for 5 different classes.
Is it possible to use the NVIDIA TAO 4 or TAO 5 to do training.

Please guide.

For example, do you mean if a dataset contains different persons and dogs, can Re-ID identify the same person across different images, also identify the same dog across different images?

Yes, like that.

Basically what i need is

  1. I want to maintain the tracking id of different objects like car, person, etc if there is some occlusion.
  2. Can i train Reid model with the same dataset which i used to train the object detection model
  3. What should be the image file name for the reid dataset for multi class? if using this (Re-Identification | NVIDIA NGC)
  4. Or if i do not use the image file name as of Market1501 can i still be able to train the model using TAO?

The ReIdentificationNet in TAO Toolkit expect data in Market-1501 format for training and evaluation.

So, it is needed to set as Market-1501 format. Refer to Data Annotation Format - NVIDIA Docs

So, there is no way to give class name in the image file name? or is it necessary to give class name in the image file name in case of training Re-ID network?

The Market-1501 dataset aims to one class person. For multi-classes, I will sync internally and let you know if it is supported.

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

After checking, currently we only support one class in TAO re-id.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.