Skip to content

Identity swaps in multi animal tracking with identity_only=True #3392

Description

@razlei25

Is there an existing issue for this?

  • I have searched the existing issues

Operating System

Linux

DeepLabCut version

3.0.0rc13
2.3.11

What engine are you using?

tensorflow

DeepLabCut mode

multi animal

Device type

NVIDIA GeForce RTX 2080 SUPER

Bug description 🐛

I have an issue with identity swaps.
Identity is set in the config file and used in the labeling (four mice with their fur dyed each in a different, clearly visible bright color) and in the tracking I set identity_only=True.
The pose estimation and assembly results are good. However, even with good raw detections, when it comes to stitching the tracklets, the identities swap. Specifically, this happens when mice go into a covered shelter and then come out of it, marked with the wrong identity. Changing tracking parameters hasn't fixed the problem.
I have 800 labeled frames from 16 videos (including many frames with full and partial occlusions caused by mice interacting and staying in the covered shelter). The setup is the same in all of the videos. The covered shelter is positioned in the same place within the setup and is the only place where the mice are not visible. It has two entrances, one on each side (this might be an additional source to the problem, since mice can enter and exit on either side randomly).

The same problem was described in issue #2867, however I was not able to avoid the problem in any DLC version I tried.
I have been using version 3.0.0rc13 (network architecture: DLCRNet_ms5; augmentation : multi-animal-imaguag) with a Tensorflow engine. After seeing the above issue, I tried version 2.3.11 (also with Tensorflow) with with no improvement (I tried both using the existing raw detections from v3 and training a new model on v2, with the same settings and shuffle, but the results were worse).

Any suggestion would help. Thank you.

Steps To Reproduce

No response

Relevant log output

Anything else?

No response

Code of Conduct

Metadata

Metadata

Labels

performanceRelated to (poor or unexpected) performance in any context (inference, training, labeling,...)reID

Type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions