Release version 1.1
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.