cnn-rnn-siamese-video-similarity

Deep CNN + LSTM siamese network for video similarity

Image-based siamese Results can be found here

Website for Visualizing Data (Only visible inside IIIT’s network)

Exp #	Conv Layer	l2-reg	batch-size	num-epochs	loss	plots	train-accuracy	val-accuracy	train-val-test split	Data-Augmentations	Runtime	learning-rate	tied-weights	convNet training	pretrained-weights
1a	conv6	0.0	64	50	contrastive	accuracy loss	9300/10167=91%	1027/1129=91%	10167-1129-0	NO	3hrs	1e-6 with a deacy of 0.85x after every epoch	YES	all layers	AmosNet weights
1b	conv6	0.0	64	50	contrastive	accuracy loss	9263/10167=91%	1024/1129=91%	10167-1129-0	NO	3hrs	1e-5 with a deacy of 0.85x after every epoch	YES	all layers	AmosNet weights
1c	fc6	0.0	64	50	contrastive	accuracy loss	9300/10167=91%	1027/1129=91%	10167-1129-0	NO	3hrs	1e-6 with a deacy of 0.85x after every epoch	YES	all layers	AmosNet weights
1d	fc7	0.0	64	50	contrastive	accuracy loss	9300/10167=91%	1027/1129=91%	10167-1129-0	NO	3hrs	1e-6 with a deacy of 0.85x after every epoch	YES	all layers	AmosNet weights
1e		0.0	64	50	contrastive	accuracy loss	9300/10167=91%	1027/1129=91%	10167-1129-0	NO	3hrs	1e-6 with a deacy of 0.85x after every epoch	YES	all layers	None

Some Jargons used above

all_layers –> all-layers in the CNN were finetuned