Improvements to Team assigner & Player tracking. #12

[tuura01]

Improvements to Team assigner & Player tracking. #12

Team Assigner:

• Collect history of past assignments for each player id, in order to majority vote on what team the player belongs to.
• Cropping of images, by some factor (can be set, by default 0.5) in order to remove information about the background to help the CLIP model with its predictions.
• Sam2 is used to create masks for the players which are then used to apply a gaussian blur to the background. I have tested various colours as well as turning the background opaque. The blur seems to be performing the best.

Player Tracking:

• Refining of bboxes via Sam2. Using the bboxes from the yolo model, create a mask for the player and then create a new, tighter bbox around the mask to further eliminate background clutter.

Pre changes:

video_1.2.mp4

Post changes:

video_1_ifstatement.mp4

[isacpaulsson]

Approved for now, sam2 provides small improvement to assignment at a major cost to inference time.
Works and looks good but is slow.