No bounding boxes

[pdhruv93]

I am creating a custom object detector for Road crack detection. The loss significantly decreases after 600 epochs. But still after that I see no bounding box on my images. Even if I use the same input image for testing, still it does not draw any bounding box. Any idea what is the issue and I have seen same issue reported multiple times: https://stackoverflow.com/questions/55830479/no-bounding-box-after-custom-traning-yolov3-model

Here is my Google colab notebook: https://colab.research.google.com/drive/16nhjP0MNIRUZv0SCGV4UsQzwEPRIn-iF#scrollTo=a7qhTxIyKdRQ

[mpky]

@pdhruv93 I'd recommend training till the loss gets down to around 1.5 and then setting the threshold quite low on the first detection attempt

[kmarrip]

@pdhruv93 loss is around 1.5 in your dataset, try getting it around 0.5.

[guias12]

How many images do you have on your dataset? I was facing the same issue the solution was to train again with more images. First time I trained with about 300 images, on the second time(when everything worked fine) I used ~1200 for each class

[kmarrip]

@guias12 have the model train on 3-5 imags first, overfit the model on these images, if the loss is around 0.5 then the learning rate is ok for the dataset, if it doesnt overfit then try chaning your learning rate. if the loss is indeed around 0.5 then try retraining the model, atleast 500 uniform images for each class is good enough for the model