基于 MMDetection 和 SlowFast 的动作时空定位,可以灵活选用 MMDetection 中任意检测模型进行动作时空定位。
先安装 MMDetection
然后安装 python 包
pip install opencv-python decord pytorchvideo ipdb
可以轻松的选用多种检测模型(参考 MMDetection 模型库) 作为检测器进行动作时空定位
使用 Deformable DETR (ICLR'2021)
export CUDA_VISIBLE_DEVICES=0
python action_det.py --video demo/run_the_city.mp4 --imsize 224 \
--config configs/deformable_detr/deformable_detr_twostage_refine_r50_16x2_50e_coco.py \
--checkpoint weights/deformable_detr_twostage_refine_r50_16x2_50e_coco_20210419_220613-9d28ab72.pth \
--out demo/run_the_city_demo.mp4
export CUDA_VISIBLE_DEVICES=0
python action_det.py --video demo/run_the_city.mp4 --imsize 224 \
--config configs/yolox/yolox_x_8x8_300e_coco.py \
--checkpoint weights/yolox_x_8x8_300e_coco_20211126_140254-1ef88d67.pth \
--out demo/run_the_city_demo.mp4
and more ...
Thanks for these wonderful works:
[1] mmdetection
[2] yolo_slowfast
[3] ZQPei/deepsort
[5] AVA: A Video Dataset of Spatio-temporally Localized Atomic Visual Actions. paper
[6] SlowFast Networks for Video Recognition. paper