Detecting wrong objects about realtimeobjectdetection HOT 4 OPEN

What's in the background images? Are there any detectable examples in these?

from realtimeobjectdetection.

What's in the background images?
Some random pics (cats, for example) that have nothing to do with my gestures.
Are there any detectable examples in these?
No. I created for this pics .xml files and deleted from there block "Object". I also changed some code in generate_tfrecord.py like in the link above.

from realtimeobjectdetection.

Hmmm, I don't think those will add much value as there aren't any labels being generated in the tfrecords file. To improve accuracy I'd suggest the following:

1. Perform image augmentation e.g. brightness, contrast, scaling and cropping
2. Train for longer until you hit a low loss metric for both train and validation
3. Use more photos for each class and under different conditions

from realtimeobjectdetection.

I added around 30 pics for each class, changed some of them(blur, brightness, contrast and etc.).
I checked detection at 0.079(28300 steps). It was good, but still sometimes detected my head as index finger and with accuracy around 50-60%. I continued training for a lot (209000 steps), stopped at 0.030 loss and results are worse than at 0.079 loss. I have some false positive detections, but, luckily, not as much as before.

Your suggestions helped a lot. Before I asked you I changed background in my pics for different colors, thought it would help somehow.

Now I will try adding more pics(100) to each class and will be stopping and checking. Did not think that at lower loss it will be performing worse.

Also what I need to do to get id of detected class in python script.

from realtimeobjectdetection.