akshaykhatri639 / ssd

Implementation of Single Shot Multibox Detector

ssd

Implementation of Single Shot Multibox Detector

Simplifications made over the paper

• Instead of selecting the top 3N boxes to include in the loss where N is the number of ground truth boxes for this sample, we select top k where k is fixed across all samples. This makes it necessary to do a hyperparameter search over k but simplifies the implementation, atleast in tensorflow

Todo:

• Find a setting of k that works well
• Implement evaluation metrics (mAP). Currently just working with precision and recall by treating it as classification over the default boxes
• Train on bigger dataset, add data augmentation
• Regress from of default box locations to make the predicted bounding boxes tighter.