Training and Inference

[abhigoku10]

Training and Inference

[wei-tim]

@abhigoku10
Hi, thanks for your interest in our project.

1. The model can be run on CPU as long as the PC has sufficient RAM.
2. Definitely. As introduced in our paper, we aim to make the backbone replaceable with different choices. For architectures with FPN (like RetinaNet) we didn't test whether they work similarly as 2D case, but I think it's a very interesting topic.
3. We trained our model with 224x224 images.
4. Our inference model would work on videos and is able to recognize actions when the performed actions do not diverge too much from the datasets. And our model is based on YOLOv2, which might be not so good at detecting tiny persons, in such cases, the results are kind of unsatisfying.

[Byronnar]

@abhigoku10
Hi, thanks for your interest in our project.

1. The model can be run on CPU as long as the PC has sufficient RAM.
2. Definitely. As introduced in our paper, we aim to make the backbone replaceable with different choices. For architectures with FPN (like RetinaNet) we didn't test whether they work similarly as 2D case, but I think it's a very interesting topic.
3. We trained our model with 224x224 images.
4. Our inference model would work on videos and is able to recognize actions when the performed actions do not diverge too much from the datasets. And our model is based on YOLOv2, which might be not so good at detecting tiny persons, in such cases, the results are kind of unsatisfying.

Thanks for your projects!
How should I do to test with a video and get the detection results?

[wei-tim]

@Byronnar

Hi,
thanks for your interest in our project.
You can follow the validation procedure and get predictions of bounding boxes as well as action labels.

[okankop]

I believe the issue is resolved!