[pic2card] Pipeline and Model Enhancements

[haridas]

• Detr model first level results are out.

• Looking good, but duplicate values are there, have to find way to remove
  it.
• Measure the mAP score against the benchmark dataset.

• Voc2coco data convertor.

• xml_to_csv and generate_tfrecords commands updated.

• image label has to be synced so that tfrecords are in correct format.
• xml_to_csv command is now more generic.
• Notebook update specific for the dataset experiments.

• Correctly set the category_ids of objects.

• Faster-RCNN updated hparams

This has been resulted into better validation and test accuracy.
Mainly the tuning has been done with aspect ratios of the image.

• Added model registry and realted pipelines.

• Abstract class for the object detection classes
• We can add multiple implementations and select on for the main
  pipeline
• Classes are structured with compossibility in mind

• Inference using libtorch for detr model.

Speed enhancments at inference time.
Helps to reduce the dependency of all python torch library stack.

• Columns grouping and image size labels

• columns grouping and image size to labels

• image size to labels logic

• removed the check of exact match of number of objects predicted

• resolved review comments

• updated the predict json test elements count

• Standardizing the object detection model pipeline.

• Helps to switch different object detection models without changing the
  pic2card pipeline.

• Both Detr and RCNN models are workig with pic2cad.

• Fixed test cases, to support multiple models.

• Latest Faster RCNN model frozen dump.

• Debug api with new bbox draw method.

Model switching are tested using env variable, and test cases are all
passing.

• Lint fixes and pep8 standardization.

Now only C901 (too complex implementation ) only exempted.

• Fail build if lint test has problems.

Co-authored-by: Keerthana Manoharan 43063410+Keerthana786@users.noreply.github.com

Microsoft Reviewers: Open in CodeFlow