Solved – SSD MobileNet v1 loss not converging bounding boxes all over the place

I've trained SSD MobileNet v2 model using Tensorflow API on my own dataset of ~4k dog pictures and it displays bounding boxes all over the place. I've trained with batch size 1.
enter image description here

The same dataset trained on faster rcnn works really well, and detects dogs properly.
I've run 50k iterations, and total loss was spiking from 1.5 to 3 all the time in the end, with random bigger spikes, but overall it went down from 20

There wasn't a smooth and gradual decline, but a really spiky and uneven one with both models

Learning graph
enter image description here

I know it's a fast model with low precision, but it should at least more or less detect the object. How can I increase accuracy so it at least partially detects objects without showing completely false bounding boxes ?

You can try adjusting the threshold in the code for inference. Default is 0.5, try out 0.7 or something like that. I have been having the same issue and it works perfectly with the faster RCNN model. Here is an insight –

Similar Posts:

Rate this post

Leave a Comment