Tuning params to improve detection and reduce false positives

[Parry26]

I have trained a fast r-CNN on my own dataset and it is working quite well (about 80% accurate for the most probable bbox). I now want to tune it to improve performance. I have run it for 100k epochs.

Manually checking (printing the images, with the bbox overlaid) I can see that in some cases, very obvious/prominent target objects are ignored. In others, there are false detections (normally objects closely resembling the target).

I would be very grateful for advice on which params to tune (presumably in the config file)?

Also, tuning the NMS and CONF thresholds (currently both 0.8).

Thanks very much.

[ranju2015]

@Parry26 Have you got any improvement? Please let me know, I am looking for the same.

[fawadhussain703]

Did you Guys Resolved this issue @Parry26 @ranju2015 ?