Accuracy problem when learning after extracting 6 classes from coco dataset(yolov3)

[yacad]

Hi AlexeyAB.

I extracted 6 classes(person, car, truck, bus, motorcycle, cycle) from coco dataset used cocoapi(The number of training images is about 70,000). And I trained as you explained. However, accuracy was lower than when training using 80 existing classes. I used multi-GPU and as you explained, I first trained on 1 GPU for 1000 iterations and then run training with multi-GPU using partially trained model /backup/yolov3_1000.weights. You explained that the max_batches value should be classes 2000, but the training were not enough, so we set classes 4000. The rest followed your instructions.

The detailed configuration file is attached here. config.zip

When I run the program using the weight file I created, I get the following result.

When I run the program using the existing 80 classes weight file, I get the following result.

As can be seen from the results, the recognition rate is lower than when training using 80 classes weight file.

Why did this result? Was there a problem with my training?

[AlexeyAB]

• Show AP values for car for default yolov3.cfg and your model

• Set max_batches=70000 steps=50000,60000 and train up to 70000 iterationns

https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

change line max_batches to (classes*2000 but not less than number of training images, but not less than number of training images and not less than 6000), f.e. max_batches=6000 if you train for 3 classes