own dataset

[styap94]

hi @harvitronix , i'm using your code to train with different dataset, my data is about sign language which involve both static and dynamic gesture. i have 10 classes, each gesture with 30 videos for training(21 videos) and testing(9 videos), my problem is when i using Method 4: lrcn, my val_error not really decreasing, always get 2.30 of val_error in the end. but when train with method 5: mlp, the best result i could get is 0.77.

My intention is to use method 4, because mlp will be very hard to run in real time (always need to extract features). So, how can i improve my result of lrcn? or there is better algorithm for my case? thank you so much for answering my doubt

[harvitronix]

Hi @hanako94. My guess is your dataset is simply much too small. Unless each video is quite long and you can break it into smaller pieces, it's going to be hard to learn sequences from just 21 data points. If there's a larger sign language dataset out there that you can train on, then use transfer learning to fine-tune to your dataset, that may be better.