MaskFormer baseline

[CJcool06]

Running a long experiment will test the validity of the current implementation and give us a standardised experimentation procedure.

Pre-requisites:

• [x] #18
• [x] Read TAO-OW

Project Plan:

1. Choose evaluation metric: Proposal Recall (background & objectness scores)
2. Decide if we are using pre-trained models: Yes
3. If not, choose training dataset: None
4. Choose evaluation dataset: TAO-OW (Validation set)
5. Choose evaluation models: MaskFormer, Mask R-CNN, Mask2Former

Experimental Plan:

1. Download and unpack TAO-OW dataset
2. Download pre-trained Mask R-CNN and MaskFormer models
3. Calculate mIOU and recall metrics for predictions
4. Generate results

What's been done:

• Re-written codebase for easier use throughout project lifespan
• Added a post-processing step to get bounding boxes from panoptic segmentation masks
• Simplified model configs
• New config options
• Simplified running by using bash scripts and make
• Inference and metric calculation happens at once
• Reduced amount of files resulting from inference (hundreds instead of hundreds of thousands)
• Easier visualisation of masks and bounding boxes

[CJcool06]

Notes:

• The box annotations in TAO are in the format of [x, y, w, h], w: width, h: height.

[CJcool06]

This issue is being closed for organisational reasons. The work outlined in this issue has been completed, however the overall project is being continued in a different issue.