Low Recall

[clydebailey]

I am training an XNOR version of tiny-yolov2 with 7th and 8th layer in XNOR. I found out some lines commented out 133-136 from the convolutional_kernels.cu file. https://github.com/AlexeyAB/darknet/blob/7f4b859514c7fcae1f9cc86276be31f133b5bc7c/src/convolutional_kernels.cu After uncommenting these lines, i trained for more than 100,000 iterations and got a decent precision but a bad a recall.

1. Why were those lines commented in the convolutional_kernel.cu file when we need the binarization input state?

2. Why am I getting a bad recall rate? precision: 0.59 TP: 107 FP:73 FN:8793

[sporterman]

@clydebailey have you met the follow trouble? hope some advice

[clydebailey]

@sporterman No. Your problems are related to your data files. Make sure that you have pointed your training data properly in obj.data and obj.names files

[sporterman]

@clydebailey i'sure the two files are correct, are there any other paramsters i need change

[clydebailey]

please show the .data and .names files along with the contents of training set.

[sporterman]

@clydebailey /database/yolov3/darknet1/data/obj.data: /database/yolov3/darknet1/data/logo.names: /database/yolov3/darknet1/data/train.txt: /database/yolov3/darknet1/data/obj: train command: ./darknet detector train data/obj.data cfg/yolo-obj.cfg darknet53.conv74

[yongcong1415]

you dataset is coco? i only trained at pascal format,

[sporterman]

@yongcong1415 no. my personal dataset, now the obj is decreasing while training,,.....

[yongcong1415]

@yongcong1415 no. my personal dataset, now the obj is decreasing while training,,.....

i mean you dataset format,you need use yolo_mark.exe to check you boxes position;

[clydebailey]

Try maintaining your images under "JPEG images" and annotations under "Annotations". This is the file format for Pascal training. And yes, use yolo_mark to check annotations.

[sporterman]

@clydebailey how to use yolo_mark to check my data and label,

[clydebailey]

1. Clone https://github.com/AlexeyAB/Yolo_mark/
2. Make sure Opencv is installed
3. Keep all your images and annotations(yolo format) in x64/Release/data/img
4. Update the .data and .names in /data
5. Run ./linux_mark.sh