Open yjh0410 opened 1 year ago
Hello! Do you plan on adding support for more resource efficient networks in the 3D backbone?
@jaca-pereira I have added 3D-ShuffleNet-v2 in my repo. I will update the performance of YOWO with efficient 3D backbone in the future.
@jaca-pereira Hi !Dear friend, I recently release the YOWO-Nano whose 3D backbone is the 3D-ShuffleNet-v2.
Hello! Thank you very much, I'll check it out.
Thanks for the open source of YOWO, a real-time method in spatio-temporal action detection task. Recently, I follow this repo. to reimplemented
YOWO
and achieve better performance, as shown in the tabels below. I name this YOWO as YOWO-Plus. We also design a efficient YOWO, YOWO-Nano whose 3D backbone is the 3D-ShuffleNet-v2-1.0x proposed by the authors of YOWO. My code is available at https://github.com/yjh0410/PyTorch_YOWO.Improvement
Better 2D backbone: We use the weights of YOLOv2 from my project. Our YOLOv2 achieves a significantly higher AP (27 AP with 416 input) on the COCO dataset.
Better label assignment: For a groundtruth, we assign the anchor boxes with IoU higher than the threshold 0.5, so each groundtruth might be assigned with multiple anchor boxes.
Better loss: We deploy GIoU loss as the box regression loss. As for the conference loss and classification loss, they are same as the ones used in YOWO. Finally, all the losses are normalized by the batch size.
Experiment