search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.77k stars 7.96k forks source link

Another problem for testing for small objects in yolov3-tiny #1340

Open RosaShao opened 6 years ago

RosaShao commented 6 years ago

Thanks for your help at first. But I got another OOM. Please help me again, thank you! I can train data when I use: set here stride= 8 and layers = -1, 4.

But after training, I can't test photo. When I tested, after 23 floor, I got that:

Total BFLOPS 30.161 CUDA Error: out of memory darknet: ./src/cuda.c:36: check_error: Assertion `0' failed. Aborted (core dumped)

But I use the same yolov3-tiny.cfg.(batch=64, subdivisions=16) I ever try to change (batch=1, subdivisions=1), (batch=32, subdivisions=8),(batch=32, subdivisions=16) or (batch=64, subdivisions=32), but still false. Could you help me? Thank you!

AlexeyAB commented 6 years ago

When I tested, after 23 floor, I got that:

  1. What is the "23 floor"?
  2. What command do you use?
RosaShao commented 6 years ago

(1)Sorry, it's after 23 layer. (2) I use: ./darknet detector test cfg/voc.data cfg/yolov3-tiny.cfg results/yolov3-tiny_27900.weights VOCdevkit/VOC2018/JPEGImages/000390.jpg

AlexeyAB commented 6 years ago
RosaShao commented 6 years ago

Did you train your ownl yolov3 for PascalVOC or for MS COCO? No, I use myself photos like the type of PascalVOC. What classes= are in the yolov3-tiny.cfg? 13 Can you run successfully yolov3-tiny with default model? https://pjreddie.com/media/files/yolov3-tiny.weights No, it fails, too. The same result: Total BFLOPS 30.161 CUDA Error: out of memory darknet: ./src/cuda.c:36: check_error: Assertion `0' failed. Aborted (core dumped)

Thank you!

RosaShao commented 6 years ago

I changed width and height to 416, then I can test. But the effect of detecting small objects is much worth than yolov3 with stride= 8 and layers = -1, 4. Could you give me another suggestion? Thank you very much!

RosaShao commented 6 years ago

I'm trying change width and height to 960, angle=0, saturation = 1.5, exposure = 1.5, hue=.1. Teating and training OK now. I'm waiting results, thank you very much!

alexanderfrey commented 5 years ago

How is detection of small objects with yolo-tiny ? Better or worse than yolov3 ?

AlexeyAB commented 5 years ago

@alexanderfrey

yolov3.cfg better than yolov3-tiny.cfg for small objects, because yolov3.cfg has a [route] layer to the layer with size 52x52, while yolov3-tiny.cfg has a [route] layer to the layerwith size 26x26 only.