SSD from scratch

The following model has been trained on google colab from scratch:

1. SSD with Multi Box Loss

2. SSD with Focal Loss

Summary

Model	epochs	Change in train loss	Loss Function	mAP	weights
SSD with Multi Box Loss	160	12.853 -> 3.0090	MultiBox Loss with Cross Entropy	58.1	Google drive
SSD with Focal Loss	175	2.5304 -> 0.803	MultiBox Loss with Focal Loss	In progress	Google drive

TODO

• Run for 300 epochs to match to research paper results.
• Compare SSD results for different Loss Functions.
• Finetune the model using VOC 2012 dataset
• Play with Batch Norm and Dropout in Auxiliary Convolution
• Multi-GPU training