Add visualization of predicted bounding boxes on Weights & Biases

[soumik12345]

🚀 Feature Request

Add feature to optionally perform inference with the model while training on a small subset of the validation data at a fixed interval and log them on Weights & Biases. This helps us gain further insight into how the model performs on specific examples from the validation set.

This feature is part of similar Weights & Biases integrations like YOLOv5, YOLOv8, Keras, etc.

• This report shows a media panel with the bounding boxes logged at the end of every training epoch for YOLOv5. This slider lets us visualize the predicted bounding boxes logged after every epoch.
• This report shows the same as an interactive wandb.Table for YOLOv8.
• This report shows the same for mmdetection as an interactive wandb.Table.

Proposed Solution

Possible ways to implement:

• Implement as a Callback by overriding the on_validation_batch_end.
• Implement as part of the WandBSGLogger.

[soumik12345]

@ofrimasad @BloodAxe
I would love to discuss this!