What is the maximum number of classes YOLOv2 can detect ?

[dnaveenr]

I have trained Tiny-YOLOV2 for more than 300 classes for custom dataset but when I test, I see no detections on the test data after more than 20k,30k iterations. I tried reducing the threshold( -thresh 0.01) but still no detections. I had tried a similar test for 100 classes and results were quite good. Is there a limit on the maximum number of classes YOLOv2 can detect ? In abstract of YOLO9000 paper, I see this :

But YOLO can detect more than just 200 classes; it predicts detections for more than 9000 different object categories. And it still runs in real-time.

Does this mean YOLOv2 can detect a maximum of 200 classes only ? Please help.

[AlexeyAB]

• What mAP can you get?
• Check your dataset using Yolo_mark
• You should train about 2000 iterations for each class, so for 300 classes = train 600 000 iterations
• Yolo v2 can detect more than 200 classes. But if you want to detect 1000-10000 classes - the better to use
  • Yolo9000: https://github.com/AlexeyAB/darknet#using-yolo9000
  • Yolo v3: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

[dnaveenr]

Thanks a lot for the quick reply. When I run ./darknet detector map command on the training dataset. I get this :

for thresh = 0.25, precision = -nan, recall = 0.00, F1-score = -nan for thresh = 0.25, TP = 0, FP = 0, FN = 6132, average IoU = -nan % mean average precision (mAP) = 0.000000, or 0.00 %

But during training, I didn't notice any nan's. The dataset seems to be in the right format, but anyway when check using Yolo_mark and get back.

Can the dataset have filenames like image- (9).jpg/txt , image - (1).jpg/txt etc, because the 300 class dataset contains filenames of this format. Would filename be an issue ? I didn't get any errors while training though.

[AlexeyAB]

mAP is bad.

Can you show your obj.data file and command for training?

Can the dataset have filenames like image- (9).jpg/txt , image - (1).jpg/txt etc,

Image filename can contain (9) But image filename can't contain images or JPEGImages or raw: https://github.com/AlexeyAB/darknet/blob/c2c8595b083ec3586a99bb913b8a986e81e3a42a/src/data.c#L299-L302

[dnaveenr]

Yes. It is bad. Okay, so filename is not an issue. Content of obj.data is :

classes = 313 train = data/train4.txt valid = data/train4.txt names = data/obj.names backup = backup/

Command used for training is :

./darknet detector train data/obj.data cfg/yolov2-tiny-obj.cfg yolov2-tiny-voc.conv.13

I have changed yolov2-tiny-obj.cfg file appropriately( i,e classes and filters ).

[AlexeyAB]

• Do you use batch=64 subdivision=8 in cfg-file?
• Did you change learning rate, burn_in or steps and scales?
• Did you check your dataset using Yolo_mark? https://github.com/AlexeyAB/Yolo_mark

[dnaveenr]

Yes, I used batch = 64 and subdivisions = 8 in the cfg file. I didn't make any changes. Just used the same parameters as in yolov2-tiny-voc.cfg

I just checked my dataset annotations. They are perfect.

Also there were few extra annotation files (.txt) files without images in the dataset folder I noticed. [ Eg : image7.txt is present, but there is no image7.jpg] I don't think this should be an issue since in train.txt we provide all image paths.

[dnaveenr]

Okay. I removed the extra annotation files in the dataset folder and retrained the model. After 1k iterations, I tested with -thresh 0.01 . I see some predictions now. So I think this was the issue.

Thanks for your help. :)

[anguoyang]

Hi@AlexeyAB, is there a performance(accuracy) comparison between yolov3 and yolo-9000 if training with custom objects? it seems that there is no training tutorial on yolo-9000

[AlexeyAB]

I didn't train yolov3 for large number of object. And I didn't compare them. I think, for classes < 100 the yolov3 much more accurately.

[anguoyang]

is it possible to add resnet/densenet to yolov3 just as in yolov2? is it necessary? thank you