Is there an existing issue for this?
Bug description
Hi! I have been training a model with the new version of DLC. In my set up, I have 4 mice that are clearly identifiable (4 different colors). When I train the model and check it against the labelled frames, performance seems quite good. However, upon creating labeled videos, I noticed huge swaps between mice idenitity. These occur especially when the mice are in close proximity (despite focusing on frames where interactions occur during the frame extraction/labelling phase). Example: if mouse A and B are approaching each other, most of the time A will be incorrectly recognized as B, and viceversa. However, there are also several instances where A and B are approaching, mouse C is on the opposite side or very far, and A will be recognized as C, B as A, and C as B. Whenever all mice are far from each other tracking is good.
For context, I have trained with 500+ labelled frames. I have tried both Pytorch and TF engines, and it leads to the same problem. This is also not improved if I try to refine tracklets.
I went back to older versions of DLC I was using (DLC 2.3.3 and 2.3.5), and training with the same frames and shuffle, gave me good results, so it seems to be a specific problem of version 3+.
Thanks ☺️
Operating System
Windows
DeepLabCut version
3.0.0rc6
DeepLabCut mode
multi animal
Device type
NVIDIA GeForce RTX 4090
Steps To Reproduce
No response
Relevant log output
Anything else?
No response
Code of Conduct
Is there an existing issue for this?
Bug description
Hi! I have been training a model with the new version of DLC. In my set up, I have 4 mice that are clearly identifiable (4 different colors). When I train the model and check it against the labelled frames, performance seems quite good. However, upon creating labeled videos, I noticed huge swaps between mice idenitity. These occur especially when the mice are in close proximity (despite focusing on frames where interactions occur during the frame extraction/labelling phase). Example: if mouse A and B are approaching each other, most of the time A will be incorrectly recognized as B, and viceversa. However, there are also several instances where A and B are approaching, mouse C is on the opposite side or very far, and A will be recognized as C, B as A, and C as B. Whenever all mice are far from each other tracking is good.
For context, I have trained with 500+ labelled frames. I have tried both Pytorch and TF engines, and it leads to the same problem. This is also not improved if I try to refine tracklets.
I went back to older versions of DLC I was using (DLC 2.3.3 and 2.3.5), and training with the same frames and shuffle, gave me good results, so it seems to be a specific problem of version 3+.
Thanks☺️
Operating System
Windows
DeepLabCut version
3.0.0rc6
DeepLabCut mode
multi animal
Device type
NVIDIA GeForce RTX 4090
Steps To Reproduce
No response
Relevant log output
Anything else?
No response
Code of Conduct