Highlights

Scenarios added

Tracking:

• Added state-of-the-art support for multi-object tracking based on the FairMOT approach described in the 2020 paper "A Simple Baseline for Multi-Object Tracking".
• Reproduced published accuracies on the popular MOT research benchmark datasets.
• Notebooks for training using a custom dataset, and for reproducing results on the MOT dataset.

Action Recognition

• Added state-of-the-art support for action recognition from video based on the R(2+1)D approach described in the 2019 paper "Large-scale weakly-supervised pre-training for video action recognition".
• Reproduced published accuracies on the popular HMDB-51 research benchmark dataset.
• Notebooks for training using a custom dataset, and for reproducing results on the HMDB-51 dataset.

Highlights

Scenarios added or expanded

Similarity:

• Implemented state-of-the-art approach for image retrieval based on the BMVC 2019 paper "Classification is a Strong Baseline for Deep Metric Learning".
• Implemented popular re-ranking approach based on the CVPR 2017 paper "Re-ranking Person Re-identification with k-reciprocal Encoding".
• Reproduced published accuracies on three popular research benchmark datasets (CARS196, CUB200, and SOP).
• Notebook added which shows how to train and evaluate the approaches on a custom dataset.

Detection:

• Added Mask-RCNN functionality to detect and segment objects.
• Added speed vs. accuracy trade-off analysis using the COCO dataset for benchmarking.
• Improved visualization of e.g. predictions, ground truth, or annotation statistics.
• Notebooks added which show how to: (i) run and train a Mask-RCNN model; (ii) evaluate on the COCO dataset; (iii) perform active learning via hard-negative sampling.

Keypoint:

• New scenario.
• Notebook added which shows: (i) how to run a pre-trained model for human pose estimation; and (ii) how to train a keypoint model on a custom dataset.

Action (in 'contrib' folder):

• New scenario.
• Added two state-of-the-art approaches for action recognition from video: (i) I3D from the famous 2017 "Quo Vadis" paper; and (ii) R(2+1)D described in the 2019 paper "Large-scale weakly-supervised pre-training for video action recognition".
• Functionality and documentation how to annotate own video data.

Scenarios

Classification:

• Introduction notebooks that include the basics of training a cutting edge classification model, how to do multi-label classification, and evaluating speed vs accuracy
• Advanced topic notebooks that include hard-negative mining, and basic exploration of parameters
• Notebooks that show how to use Azure ML to operationalize your model, and Azure ML Hyperdrive to perform exhaustive testing on your model

Similarity:

• Introduction notebooks that performs basic training and evaluation for image similarity
• Notebooks that show how to use Azure ML hyperdrive to perform exhaustive testing on your model

Detection:

• Introduction notebooks that performs basic training and evaluation for object detection
• Notebooks that show how to use Azure ML hyperdrive to perform exhaustive testing on your model