Using Deep Convolutional LSTM Networks for Learning Spatiotemporal Features.

This paper explores the use of convolutional LSTMs to simultaneously learn spatial- and temporal-information in videos. A deep network of convolutional LSTMs allows the model to access the entire range of temporal information at all spatial scales. We describe our experiments involving convolutional LSTMs for lipreading that demonstrate the model is capable of selectively choosing which spatiotemporal scales are most relevant for a particular dataset. The proposed deep architecture holds promise in other applications where spatiotemporal features play a vital role without having to specifically cater the design of the network for the particular spatiotemporal features existent within the problem. Our model has comparable performance with the current state of the art achieving 83.4% on the Lip Reading in the Wild (LRW) dataset. Additional experiments indicate convolutional LSTMs may be particularly data hungry considering the large performance increases when fine-tuning on LRW after pretraining on larger datasets like LRS2 (85.2%) and LRS3-TED (87.1%). However, a sensitivity analysis providing insight on the relevant spatiotemporal temporal features allows certain convolutional LSTM layers to be replaced with 2D convolutions decreasing computational cost without performance degradation indicating their usefulness in accelerating the architecture design process when approaching new problems.