Adding a new task to Hub: object tracking in videos

[kadirnar]

[merveenoyan]

Wanted to open the discussion from the discord, @osanseviero what is required for us to have a separate pipeline/task for object tracking in videos? Pinging @sgugger too!

[sgugger]

To have a pipeline in Transformers, we'd need pretrained models that do this. I don't think that is the case right now.

[kadirnar]

To have a pipeline in Transformers, we'd need pretrained models that do this. I don't think that is the case right now.

Models:

• https://huggingface.co/kadirnar/osnet_x0_5_imagenet
• https://huggingface.co/kadirnar/osnet_x0_25_imagenet
• https://huggingface.co/kadirnar/osnet_x1_0_imagenet

Supported tracking algorithms:

• [X] Sort
• [X] StrongSort
• [X] ByteTrack
• [X] OcSort
• [X] Norfair

Demo: https://huggingface.co/spaces/kadirnar/torchyolo

[osanseviero]

Hey all! Let me copy-paste the template for tracking new tasks

Note that you're not expected to do all of the following steps. This helps track all the steps required to get a new task fully supported in the Hub 🔥

• [ ] Integration with Inference API. Select at least one of the following:
  • [ ] Added a transformers pipeline
  • [ ] Added to Community Inference API for 3rd party library
  • [ ] Added to Community Inference API for generic
• [ ] Added basic UI elements (icon, order specification, etc)
• [ ] Added a widget

Integration guide: https://hf.co/docs/hub/models-tasks

[merveenoyan]

@osanseviero @sgugger I was thinking more for whether there should be a separate task for this or if this could be covered under object detection as is in our ecosystem, that's why I asked above question 🙂

[osanseviero]

I think it's different, as one task operates with video as inputs (so temporal information) while the other just operates with static input images.