馃殌 Feature
I'd like to request new tasks for:
- Video object detection (models which use multi-frame information to improve detection quality compared to single-frame detection)
- Video multi-object tracking (using either the tracking-by-detection or the joint-detection-association paradigm)
Motivation
These are two very common tasks, but I can't seem to find implementations for either of them that are part of a larger, multi-task vision library, especially with support for pytorch lightning. The closest is probably the mmtracking repository, though it lacks many convieniences of PL. If you want to do either of these tasks right now, your best bet is probably to use a single-model library like ByteTrack, CenterTrack, etc. that often are not robust in terms of training features, distributed training, etc.
Pitch
There's several models available for each. I'd maybe suggest ByteTrack for the MOT task and Temporal RoI Align for video object detection task, though any models that strike a good balance between ease of integration and performance would be great additions. Ideally this would be compatible with pytorch lightning.
馃殌 Feature
I'd like to request new tasks for:
Motivation
These are two very common tasks, but I can't seem to find implementations for either of them that are part of a larger, multi-task vision library, especially with support for pytorch lightning. The closest is probably the mmtracking repository, though it lacks many convieniences of PL. If you want to do either of these tasks right now, your best bet is probably to use a single-model library like ByteTrack, CenterTrack, etc. that often are not robust in terms of training features, distributed training, etc.
Pitch
There's several models available for each. I'd maybe suggest ByteTrack for the MOT task and Temporal RoI Align for video object detection task, though any models that strike a good balance between ease of integration and performance would be great additions. Ideally this would be compatible with pytorch lightning.