Video object tracking
- MOT 17 dataset https://motchallenge.net/data/MOT17/
-
Faster RCNN MOT17 detection benchmark https://motchallenge.net/results/MOT17Det/
S. Ren, K. He, R. Girshick, J. Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NIPS, 2015
| AP | MODA | MODP | FAF | TP | FP | FN | Precision | Recall |
|---|---|---|---|---|---|---|---|---|
| 0.72 | 68.5 | 78.0 | 1.7 | 88,601 | 10,081 | 25,963 | 89.8 | 77.3 |
-
No multi-gpu training support, only makes use of single gpu
-
Sagemaker trainining makes use of SPOT instances, need to implement checkpointing to resume training when interrupted
-
To run on command line, using the mot17 dataset
export PYTHONPATH=./src python ./src/experiment_train.py --dataset Mot17DetectionFactory --traindir ./tests/data/clips --valdir tests/data/clips --batchsize 8 --commit_id 763b78c085244fa2fe816f48545cdb520e037b51 --epochs 2 --learning_rate 0.0001 --log-level INFO --model FasterRcnnFactory --momentum 0.9 --patience 20 --weight_decay 5e-05 -
To run on SageMaker, see notebook Sagemaker.ipynb