Object detection
VISHWANATHAN007 opened this issue · comments
Hi,
I am currently training a model on object detection using ssd mobilenet v1.
I had some doubts regarding the hard negative mining and the ssd_random_crop
- In hard_negative_mining, what does min_negatives_per_image indicate?
I have generated the tf records with negative samples(containing no objects)
Does min_negatives_per_image indicate how many of these negative samples would be
considered for training? - Is ssd_random_crop data augmentation skipped for negative data in the tf record(i.e, data with no
bounding box)?
Thanks in advance :)