Problems of Speed and MR

[XiongZhongxia]

Hi, I have some questions about your works, including CFR, GAFF and MD.

Firstly, according to your paper, GAFF runs 9.34 ms on 1080TI and 11.5ms on TX2, with feature fusion. In addition, in, your latest WACV 2022 paper, Low-Cost Multispectral Scene Analysis With Modality Distillation, speed of pure Resnet18+Retinanet with 640×512 input is 7ms.

However, I tried to reproduce Resnet18+Retinanet and VGG16+Retinanet, without feature fusion but with the same input size, the speeds are 18.3 ms and 35.6 ms, respectively. Note these speeds are tested on TITAN Xp, which is a GPU with even a little bit stronger power than 1080TI.

So my question is how did you test your models and report their speeds?

In addition, I think it's very hard to deploy a model from PC GPU such as 1080TI to edge device like TX2 with just a little loss of the speed, even if with FP16 or INT8 quantization.

Moreover, why the MR in your work MD inconsistent with your previous work GAFF, since the teacher network is exactly Resnet18+Retinanet+GAFF?

I think it would be helpful if you can release your entire code and model instead of results files. Thanks.