YOLOv3 Training but not Detecting

[JoltonADsouza]

I am training on a dataset(Caltech Pedestrian) with ~90k images(640 x 640) with just a single class(person), running kmeans to find its own anchors, making required changes in filters in the .cfg file and changing class number/names in the .data/.classes files and run for an iteration of 18k with the following status:

So, the training starts at ~6000 and learns till I achieve a loss of ~0.7 and I stop training. But when I am trying to test it out on a sample video, it does not detect any person. This is what my .cfg has,

batch=64 subdivisions=16 width=640 height=640 channels=3 momentum=0.9 decay=0.0005 angle=0 saturation = 1.5 exposure = 1.5 hue=.1

learning_rate=0.001 burn_in=500 max_batches=40000 policy=steps steps=30000 scales=.1

What is exactly wrong?

[LukeAI]

can you share the chart.png ?

[JoltonADsouza]

can you share the chart.png ?

I don't have the chart.png for it. I used the https://github.com/pjreddie/darknet repo

[varghesealex90]

Hello @JasonACDsouza ,

Not sure if this will help you. When you test can you edit the cfg

From:

Testing

batch=1

subdivisions=1

Training

batch=64 subdivisions=16

TO

Testing

batch=1 subdivisions=1

Training

batch=64

subdivisions=16

[jfangah]

Did you train your model on coco dataset? You may need to check your category_id in your json file.