search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.65k stars 7.96k forks source link

Have you read about this Xailient’s Detectum? (up to 300 FPS on a Raspberry) #4305

Open isra60 opened 4 years ago

isra60 commented 4 years ago

I 've found this article and I am impressed with the FPS and high accuracy they announce.

https://www.xailient.com/post/challenges-of-running-deep-learning-computer-vision-on-computationally-limited-devices

LukeAI commented 4 years ago

I don't believe it. no details, no code. just hype.

AlexeyAB commented 4 years ago

What Dataset/Challenge did they use to get these mAP results?

1_2BqKcRZsvRUPqEOpmpilzQ

I think even MixNet https://github.com/AlexeyAB/darknet/issues/4203#issuecomment-553129025 + FPN/TridentNet + [yolo]-GIoU-layers outperform Xailient’s Detectum in speed/accuracy, if @dkurt will add squeze-n-excitation blocks for Darknet to OpenCV-dnn.

But XNOR-net (MixNet-M-Xnor + SVR) can be much faster on low-end CPUs or FPGA, especially with Yolov3-GIoU layers: https://github.com/AlexeyAB/darknet/issues/3054

https://www.researchgate.net/publication/323375650_A_Lightweight_YOLOv2_A_Binarized_CNN_with_A_Parallel_Support_Vector_Regression_for_an_FPGA

image

dkurt commented 4 years ago

I think even MixNet #4203 (comment) + FPN/TridentNet + [yolo]-GIoU-layers outperform Xailient’s Detectum in speed/accuracy, if @dkurt will add squeze-n-excitation blocks for Darknet to OpenCV-dnn.

Feel free to open an issue to not to miss this feature request. Thanks!

MichaelCong commented 4 years ago

Xailient’s Detectum? (up to 300 FPS on a Raspberry) how to get the paper?thank you

haviduck commented 4 years ago

That whole article was Just air..

AlexeyAB commented 4 years ago

@dkurt Hi,

I added Feature request: https://github.com/opencv/opencv/issues/15987

After these improvements are implemented, I will add a request for larger changes for supporting MixNet + EfficientDet + Gaussian_yolo ...

It is interesting to see how efficiently the EfficientNet networks (mainly grouped/depthwise convolutions) can be processed on Intel CPU and Intel Myriad X neurochips.

abhiTronix commented 4 years ago

Xailient has proven the Detectum software performs CV 98.7% more efficiently without losing accuracy. Detectum object detection, which performs both localization and classification of objects in images and video, has been demonstrated to outperform the industry-leading YOLOv3. Xailient achieved the same accuracy 76x faster than the Cloud Baseline, and was 8x faster than the Edge Baseline without the accuracy penalty.

Just claims.