cnn-rnn-siamese-video-similarity

Deep CNN + LSTM siamese network for video similarity

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

Following parameters were fixed

Exp # images train-accuracy val-accuracy positive-negative weather-type in pair of videos trajectory-type learning-rate convNet training features from LSTM Num LSTM Layers train-val ratio
1a accuracy loss 707/707=100% 77/77=100% 375/404 same-weather same-direction 1e-5 NO CELL-STATE 1 9:1
2a accuracy loss 1011/1011=100% 111/112=100% 443/680 all-weather same-direction 1e-5 NO CELL-STATE 1 9:1
2b accuracy loss 782/1011=77.3% 83/112=74% 443/680 all-weather same-direction 1e-5 NO OUTPUT(LAST Frame) 1 9:1
2c accuracy loss 782/1011=77.3% 86/112=77% 443/680 all-weather same-direction 1e-4 NO OUTPUT(LAST Frame) 1 9:1
2d accuracy loss 733/1011=72.5% 84/112=75% 443/680 all-weather same-direction 1e-5 YES OUTPUT(LAST Frame) 1 9:1
2e accuracy loss 791/1011=78.2% 93/112=83% 443/680 all-weather same-direction 1e-5 YES OUTPUT(LAST Frame) 3 9:1
3a accuracy loss 808/813=100% 89/89=100% 498/404 same-weather same/overlap-directions 1e-5 NO CELL-STATE 1 9:1
4a accuracy loss 932/937=99% 101/103=98% 636/404 all-weather same/overlap-directions 1e-5 NO CELL-STATE 1 9:1
4b accuracy loss 520/520=100% 520/520=100% 636/404 all-weather same/overlap-directions 1e-5 NO CELL-STATE 1 1:1
4c accuracy loss 100% 55% 636/404 all-weather same/overlap-directions 1e-5 NO CELL-STATE 1 1:9
4d accuracy loss 260/260=100% 535/780=69% 636/404 all-weather same/overlap-directions 1e-5 NO CELL-STATE 1 1:3
4e accuracy loss 417/417=100% 595/623=95% 636/404 all-weather same/overlap-directions 1e-5 NO CELL-STATE 1 3:2
4f accuracy loss 931/937=99% 101/103=98% 636/404 all-weather same/overlap-directions 1e-5 NO HIDDEN-STATE 1 9:1
5a accuracy loss 501/501=100% 55/55=100% 152/404 same-weather inverse-direction 1e-5 NO CELL-STATE 1 9:1
6a accuracy loss 502/502=100% 52/55=95% 153/404 all-weather inverse-direction 1e-5 NO CELL-STATE 1 9:1
7a accuracy loss 612/612=100% 67/67=100% 275/404 same-weather inverse/overlap-direction 1e-5 NO CELL-STATE 1 9:1
8a accuracy loss 635/635=100% 70/70=95% 301/404 all-weather inverse/overlap-direction 1e-5 NO CELL-STATE 1 9:1
9a accuracy loss 1581/1620 180/180=100% 1120/680 all-weather all-directions 1e-5 NO CELL-STATE 1 9:1
9b accuracy loss 1600/1620=100% 180/180=100% 1120/680 all-weather all-directions 1e-5 NO HIDDEN-STATE 1 9:1
9c accuracy loss 900/900=100% 859/900=95% 1120/680 all-weather all-directions 1e-5 NO HIDDEN-STATE 1 1:1
9d accuracy loss 450/900=50% 360/900=40% 1120/680 all-weather all-directions 1e-5 NO OUTPUT(LAST Frame) - LSTM WITH PEEPHOLES 1 1:1
9e accuracy loss 516/900=58% 436/900=48% 1120/680 all-weather all-directions 1e-5 NO GRU-Output(Last Frame) 1 1:1
9f accuracy loss 869/900=99% 898/900=100% 1120/680 all-weather all-directions 1e-5 NO GRU-Hidden 1 1:1