SSD: Single Shot MultiBox Detector
Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg. ECCV 2016.
Models
Using Dataset
Implementation | Accuracy | Weights | Memory | Conv Ops | etc | Link |
---|---|---|---|---|---|---|
Keras | ||||||
TensorFlow | ||||||
PyTorch |
Tip & Trick
name | for What | reference |
---|---|---|
Multi-scale featuremap | allow prediction at multiple scales | - |
Default boxes | allow separately detect objects with different ratio | 'Anchor' in Faster R-CNN |
jaccard overlay | to get best box | a.k.a. IOU |
Smooth L1 loss | loss function for bbox regression | image, paper |
mAP | eval metric | explanation |
hard negative mining | to sample from a lot of negative boxes | paper |
a trous algorithm | filling the 'holes' while transforming FC weights into CONV weight | wikipedia |
NMS(Nom-Max Suppression | to leave only one bbox (with highest confidence level) during inference time | coursera - deep learning |