Use yolov4 train - it shows Wrong annotation

[brianng0305]

Sir,

System: darknet, yolov4

Concrening with the darknet labels export, I set the Names.file as obj.names and export labels. I used the exported labels files for the yolov4 training, after loaded the model, it show error as below.

[yolo] params: iou loss: ciou (4), iou_norm: 0.07, cls_norm: 1.00, scale_x_y: 1.05 nms_kind: greedynms (1), beta = 0.600000 Total BFLOPS 59.636 avg_outputs = 491093 Allocate additional workspace_size = 52.43 MB Loading weights from /home/brian/yolo/layout/yolov4.conv.137... seen 64, trained: 0 K-images (0 Kilo-batches_64) Done! Loaded 137 layers from weights-file Learning Rate: 0.001, Momentum: 0.949, Decay: 0.0005 If error occurs - run training with flag: -dont_show Resizing, random_coef = 1.40

608 x 608 Create 6 permanent cpu-threads

Wrong annotation: x = 0, y = 0, < 0 or > 1, file: /home/brian/yolo/layout/output/train/1.txt . . . Wrong annotation: x = 0, y = 0, < 0 or > 1, file: /home/brian/yolo/layout/output/train/2.txt

Wrong annotation: x = 0, y = 0, < 0 or > 1, file: /home/brian/yolo/layout/output/train/3.txt

Wrong annotation: x = 0, y = 0, < 0 or > 1, file: /home/brian/yolo/layout/output/train/4.txt

Wrong annotation: x = 0, y = 0, < 0 or > 1, file: /home/brian/yolo/layout/output/train/5.txt

Wrong annotation: x = 0, y = 0, < 0 or > 1, file: /home/brian/yolo/layout/output/train/6.txt

So I stopped the training. why the annotation are wrong?

[jveitchmichaelis]

Are all your labels zero or only some?

[brianng0305]

All labels zero

[jveitchmichaelis]

I can't really give you any help without seeing either your dataset, or the label database you're using. If all your labels are zero then something odd is happening. In the past we've seen issues where one or two labels were placed at the origin as a bug, but never the whole set.

Can you paste an example exported label file?

[brianng0305]

7 0 0 0 0 7 0.784314 0.0560928 0.0896359 0.065764 7 0.0504202 0.0580271 0.0756303 0.0580271 5 0 0 0 0 5 0.798319 0.628627 0.137255 0.0773694 0 0 0 0 0 8 0 0 0 0 8 0 0 0 0 8 0.918768 0.804642 0.0980392 0.0812379 4 0 0 0 0 4 0.913165 0.93617 0.0868347 0.065764 6 0 0 0 0 6 0.535014 0.943907 0.117647 0.0676983 7 0 0 0 0 7 0.52381 0.804642 0.092437 0.0851064 1 0 0 0 0 1 0.154062 0.663443 0.114846 0.0870406 8 0 0 0 0 8 0.162465 0.808511 0.123249 0.0909091

[jveitchmichaelis]

Ok, I suspect this is a bug. I can probably add a check to ignore labels with zero coords. You're building from source I guess, so if I push something quickly you should just be able to rebuild.

Did you do something like tracking? It's odd that you have lots of duplicated labels with zero coords in them.

This should be handled by this, by the way: https://github.com/jveitchmichaelis/deeplabel/commit/ecc24c5d6c36818f6ab29c4817a4b045dc25c5c8

Can check one of your images - does deeplabel show more bounding boxes than you have labelled? For example you label 4 things and it shows 8.

[brianng0305]

Yes. I did tracking.(since I am not familiar with the software)

[jveitchmichaelis]

Ok, and you're working on a video sequence? You shouldn't be using tracking for other datasets.

[brianng0305]

Can check one of your images - does deeplabel show more bounding boxes than you have labelled? For example you label 4 things and it shows 8. <- yes

[brianng0305]

Ok, and you're working on a video sequence? You shouldn't be using tracking for other datasets. <- Got it. Thank you very much

[jveitchmichaelis]

Just to check, do you have the latest version of the code? e.g. run git pull

It's odd that you get zero area bounding boxes, because that error should be caught (i.e. they should never be saved in the first place).

I can add a very simple check in the exporter to skip these boxes, but that doesn't solve the root of the problem. For now just add:

            // Ignore labels with zero area
            if(label.rect.width()*label.rect.height() <= 0 ){
                continue;
            }

In darknetexporter.cpp around line 50. e.g.

  // Delete existing files for simplicity.
    if (f.open(QIODevice::WriteOnly | QIODevice::Truncate)) {
        BoundingBox label;
        foreach(label, labels){
            QString text;

            // Check if this label exists in the database
            if(id_map.find(label.classname.toLower()) == id_map.end()){
                qDebug() << "Couldn't find this label in the names file: " << label.classname.toLower();
                continue;
            }

            // Ignore labels with zero area
            if(label.rect.width()*label.rect.height() <= 0 ){
                continue;
            }

         .....

[brianng0305]

the annotations files are correct now but it shows error:

libpng warning: iCCP: known incorrect sRGB profile libpng warning: iCCP: known incorrect sRGB profile . . .

I may git clone this and re-build the files. (I am using the deeplabel 0.12 version)

[jveitchmichaelis]

Yes, I'd recommend building from master.

Not sure about the pngs. Deeplabel just copies images on export. Probably you can ignore it? Is that a warning in darknet or deeplabel?

[brianng0305]

Yes. It highly propobly is the warning from darknet. I may use yolov3 model to try. It seems better.

[jveitchmichaelis]

Can you build from master to check if that solves your label issue? Then I can close this, thanks!

[brianng0305]

Not works. Numbers of bounding boxes and lebals are not matched. I still need to add your code manually for the 'no zero' labels export.

[jveitchmichaelis]

@brianng0305 I've just pushed v0.13 (source only), but give that a go.

Your old project will still have zero-area labels. Deeplabel doesn't (at the moment) have the ability to "clean up" a project. This is just a workaround so that these labels are ignored during load/export. If you start a new project from scratch, these labels won't be saved in the first place (I hope).

https://github.com/jveitchmichaelis/deeplabel/commit/757ecc4259574e886954151a812de2fbe6f4e205

[brianng0305]

Sir,

Yes. The v0.13 works perfectly and yolo is training the data. Thank you very much.

[jveitchmichaelis]

Awesome, I hope your model trains well! Thanks for being patient.