SLR

isolated & continuous sign language recognition using CNN+LSTM/3D CNN/GCN/Encoder-Decoder

Requirements

• Download and extract CSL Dataset
• Download and install PyTorch

Isolated Sign Language Recognition

CNN+LSTM

1. four layers of Conv2d + one layer of LSTM

   Dataset	Classes	Samples	Best Test Acc	Best Test Loss
   CSL_Isolated	100	25,000	82.08%	0.734426
   CSL_Isolated	500	125,000	71.71%	1.332122

2. ResNet + one layer of LSTM

   Dataset	Classes	Samples	Best Test Acc	Best Test Loss
   CSL_Isolated	100	25,000	93.54%	0.245582
   CSL_Isolated	500	125,000	83.17%	0.748759

3D CNN

1. three layers of Conv3d

   Dataset	Classes	Samples	Best Test Acc	Best Test Loss
   CSL_Isolated	100	25,000	58.86%	1.560049
   CSL_Isolated	500	125,000	45.07%	2.255563

2. 3D ResNet

   Method	Dataset	Classes	Samples	Best Test Acc	Best Test Loss
   ResNet18	CSL_Isolated	100	25,000	93.30%	0.246169
   ResNet18	CSL_Isolated	500	125,000	79.42%	0.800490
   ResNet34	CSL_Isolated	100	25,000	94.78%	0.207592
   ResNet34	CSL_Isolated	500	125,000	81.61%	0.750424
   ResNet50	CSL_Isolated	100	25,000	94.36%	0.232631
   ResNet50	CSL_Isolated	500	125,000	83.15%	0.803212
   ResNet101	CSL_Isolated	100	25,000	95.26%	0.205430
   ResNet101	CSL_Isolated	500	125,000	83.18%	0.751727

3. ResNet (2+1)D

   Dataset	Classes	Samples	Best Test Acc	Best Test Loss
   CSL_Isolated	100	25,000	98.68%	0.043099
   CSL_Isolated	500	125,000	94.85%	0.234880

GCN

Skeleton+LSTM

Continuous Sign Language Recognition

Encoder-Decoder

Encoder is ResNet18+LSTM, and Decoder is LSTM

References

• Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet?

• Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition

• A Closer Look at Spatiotemporal Convolutions for Action Recognition

• SIGN LANGUAGE RECOGNITION WITH LONG SHORT-TERM MEMORY

• https://github.com/HHTseng/video-classification

• https://github.com/kenshohara/3D-ResNets-PyTorch

• https://github.com/bentrevett/pytorch-seq2seq