search

chanchanchan97 / ICAFusion

ICAFusion: Iterative Cross-Attention Guided Feature Fusion for Multispectral Object Detection, Pattern Recognition
GNU Affero General Public License v3.0
94 stars 7 forks source link

Inference speed #16

Open XiongZhongxia opened 4 months ago

XiongZhongxia commented 4 months ago

Hi, Jifeng. Thanks for your work!

According to your paper, the inference speed is 38 FPS，with input size=640*512. Also, after reading your code, I assume you use yolov5l as the basic model.

On the other hand, this repo is greatly benefited from the code of CFT. So, since you didn't upload the checkpoints, I tried to test CFT with yolov5l architecture and 640*512，but the speed is only 20 FPS on 3090 GPU.

I think it's hard to understand these results, given the fact that the computational complexity of your DMFF is not significantly less than CFT, which is also reported in page 13 of your paper.

So, it would be much helpful if you can provide your checkpoint and demo code.

XiongZhongxia commented 4 months ago

I just found out you uploaded your checkpoint 1 hour ago. LOL

I would like to test it immediately.

chanchanchan97 commented 4 months ago

Thank you for your attention. We have upload the checkpoint on the kaist dataset, and update our demo code. You can validate the inference speed with detect_twostream.py.

CarryHJR commented 4 months ago

using detect_twostream.py in one 3090, i got

Results saved to runs/detect/exp
Done. (45.605s)
Average Speed: 50.507205Hz

looks good

jianguo520 commented 4 months ago

using detect_twostream.py in one 3090, i got

Results saved to runs/detect/exp
Done. (45.605s)
Average Speed: 50.507205Hz

looks good How did you achieve that? I tried the latest detect_twostream.py on 4 TITAN X，i got 1710230487990

CarryHJR commented 4 months ago

your TITAN X problem

jianguo520 commented 4 months ago

Thank you for your response. I'm wondering, is the difference in graphics cards really that significant? Have you tried it on the LLVIP dataset? My experimental results on LLVIP were very poor, with a map50 of only 84.

CarryHJR commented 4 months ago

i just conduct some experiments on FLIP, using one gpu is enough, training is 1~2 hours

jianguo520 commented 4 months ago

感谢您的回复~

XiongZhongxia commented 4 months ago

using detect_twostream.py in one 3090, i got

Results saved to runs/detect/exp
Done. (45.605s)
Average Speed: 50.507205Hz

looks good How did you achieve that? I tried the latest detect_twostream.py on 4 TITAN X，i got 1710230487990

TITAN X

I also tested the inference speed for this repo on TITAN X， got nearly 19-21 FPS. GPU indeed matters.

wenhongwu commented 1 month ago

i just conduct some experiments on FLIP, using one gpu is enough, training is 1~2 hours

how much mAP have you achieved in your training