Comparison of Spatiotemporal Networks for Learning Video Related Tasks
This work provides insights for researchers in video analysis by comparing network architectures on a controlled dataset, but it is incremental as it focuses on empirical analysis without introducing new methods.
The paper tackled the problem of comparing spatiotemporal networks for video tasks by constructing a controlled MNIST-based video dataset with parameters for classification, ordering, and speed estimation, and found that models differ significantly based on task and architecture choices, with design decisions heavily impacting feature learning.
Many methods for learning from video sequences involve temporally processing 2D CNN features from the individual frames or directly utilizing 3D convolutions within high-performing 2D CNN architectures. The focus typically remains on how to incorporate the temporal processing within an already stable spatial architecture. This work constructs an MNIST-based video dataset with parameters controlling relevant facets of common video-related tasks: classification, ordering, and speed estimation. Models trained on this dataset are shown to differ in key ways depending on the task and their use of 2D convolutions, 3D convolutions, or convolutional LSTMs. An empirical analysis indicates a complex, interdependent relationship between the spatial and temporal dimensions with design choices having a large impact on a network's ability to learn the appropriate spatiotemporal features.