qengineering / yolov4-ncnn-raspberry-pi-4 Goto Github PK

When "not" using YOLOV4_TINY program generates a segmentation fault in

static int detect_yolov4(const....
.
.
objects.push_back(object); <--

Reproducable under Ubuntu as well

I cloned the repository and ran the YoloV4.cbp file, but I am getting a "Segmentation fault" error.
I am using raspberry pi 4 with code blocks

Hi I have 2 custom trained scaled YoloV4 models (YoloV4-csp & Darknet YoloV4-Tiny) to detect 10 classes. I have reformatted them to work with your examples, for csp version (Pytorch->ONNX->MNN) and TINY (Darknet->tflite & Darknet->ncnn). I'm having a hard-time figuring out how to adapt your yolov4.cpp code to set confidence or IOU thresholds for my Tiny model. I'm not going get into the YoloV4-csp model here because I couldn't convert it to ncnn and the mnn/onnx versions are another can of worms. Don't even know where to begin...

Anyways, I couldn't re-work your SSD code to get the tflite tiny model to run, but with some tweaking I did manage to get the ncnn version going with webcam infeed. I compared the topology of this repo's Tiny.param with my custom trained one and the only thing that stands out are the csp feature shortcuts. For example, the topology of my model's last few layers looks like this:

Split 27_192_bn_leaky_split 1 2 27_192_bn_leaky 27_192_bn_leaky_split_0 27_192_bn_leaky_split_1 -23330=8,3,20,20,256,3,20,20,256
Convolution 28_200 1 1 27_192_bn_leaky_split_0 28_200_bn_leaky -23330=4,3,20,20,512 0=512 1=3 4=1 5=1 6=1179648 9=2 -23310=1,1.0e-01
Convolution 29_208 1 1 28_200_bn_leaky 29_208 -23330=4,3,20,20,45 0=45 1=1 5=1 6=23040
Yolov3DetectionOutput detection_out 1 1 29_208 yolo0 -23330=4,2,6,1,1 0=10 1=3 2=2.5e-01 -23304=12,1.0e+01,1.4+01,2.3e+01,2.7e+01,3.7e+01,5.8e+01,8.1e+01,8.2e+01,1.35e+02,1.69e+02,3.44e+02,3.190e+02 -23305=3,1077936128,1082130432,1084227584 -23306=2,3.36e+01,3.36e+01
Convolution 32_236 1 1 27_192_bn_leaky_split_1 32_236_bn_leaky -23330=4,3,20,20,128 0=128 1=1 5=1 6=32768 9=2 -23310=1,1.0e-01
Interp 33_244 1 1 32_236_bn_leaky 33_244 -23330=4,3,40,40,128 0=1 1=2.0e+00 2=2.0e+00
Concat 34_247 2 1 33_244 23_167_bn_leaky_split_1 34_247 -23330=4,3,40,40,384
Convolution 35_250 1 1 34_247 35_250_bn_leaky -23330=4,3,40,40,256 0=256 1=3 4=1 5=1 6=884736 9=2 -23310=1,1.0e-01
Convolution 36_258 1 1 35_250_bn_leaky 36_258 -23330=4,3,40,40,45 0=45 1=1 5=1 6=11520
Yolov3DetectionOutput detection_out 1 1 36_258 yolo1 -23330=4,2,6,38,1 0=10 1=3 2=2.5e-01 -23304=12,1.0e+01,1.4e+01,2.3e+01,2.7e+01,3.7e+01,5.8e+01,8.1e+01,8.2e+01,1.35e+02,1.69e+02,3.44e+02,3.19e+02 -23305=3,1065353216,1073741824,1077936128 -23306=2,1.68e+01,1.68e+01

In the yolov4.cpp code, when I set ex.extract("yolo0", out), the confidence of my detected object is ~80% where as if I use the last layer's output ("yolo1") I get ~20-30% for the same object. I can't figure out how I should be merging the results of these two parallel output layers or how to set their detection thresholds like you do in the YoloV5 example. Any advice with code would be greatly appreciated!