Interpreting Video Features: A Comparison of 3D Convolutional Networks and Convolutional LSTM Networks

Abstract

A number of techniques for interpretability have been presented for deep learning in computer vision, typically with the goal of understanding what the networks have actually learned underneath a given classification decision. However, interpretability for deep video architectures is still in its infancy and we do not yet have a clear concept of how to decode spatiotemporal features. In this paper, we present a study comparing how 3D convolutional networks and convolutional LSTM networks learn features across temporally dependent frames. This is the first comparison of two video models that both convolve to learn spatial features but that have principally different methods of modeling time. Additionally, we extend the concept of meaningful perturbation introduced by Fong & Vedaldi (2017) to the temporal dimension to search for the most meaningful part of a sequence for a classification decision. (a) Moving something and something away from each other.

Play Fair: Frame Attributions in Video Models

“…Experimental Setup. Our experiments are conducted on the validation set of the large-scale Something-Something v2 dataset [10], frequently used to probe video models [1,29,33] due to its fine-grained nature, large number of classes (174), and the temporal characteristics of most of its classes e.g. "Pulling [...] from behind of [...]".…”

“…Evaluating a model on feature subsets is challenging as models rarely support the notion of a 'missing feature'. Two approaches exist: re-training the model on all combinations of features [28] or substituting missing features with those from a reference [14,16,22,24,25,29,30], but both approaches have limitations. Retraining is computationally infeasible for more than a handful of features, and the choice of reference in feature substitution has a significant impact on the resulting attribution values [18,31].…”

“…Explaining Video Models. Most works [29,[35][36][37][38]41] in attribution for video understanding use backprop methods originally designed for images, such as Grad-CAM [20] or EBP [21]. A few recent works propose video-specific attribution methods.…”

“…They extend the approach in [48] by backpropagating gradients to different depths in the network. Mänttäri et al [29] apply meaningful perturbations [24] to learn a temporal mask over the input. To keep the number of input frames fixed, they replace missing frames by duplication.…”

Pulmonary COVID-19: Learning Spatiotemporal Features Combining CNN and LSTM Networks for Lung Ultrasound Video Classification

“…This class of RNN favors work with long sequences. This is why LSTMs have been applied to different areas of knowledge, including medical diagnosis, being frequently used in problems involving video classification and recognition of human activities [ 92 , 93 , 94 , 95 ].…”