For egocentric vision tasks such as action recognition, there is a relative scarcity of labeled data. This increases the risk of overfitting during training. In this paper, we address this issue by introducing a multitask learning scheme that employs related tasks as well as related datasets in the training process. Related tasks are indicative of the performed action, such as the presence of objects and the position of the hands. By including related tasks as additional outputs to be optimized, action recognition performance typically increases because the network focuses on relevant aspects in the video. Still, the training data is limited to a single dataset because the set of action labels usually differs across datasets. To mitigate this issue, we extend the multitask paradigm to include datasets with different label sets. During training, we effectively mix batches with samples from multiple datasets. Our experiments on egocentric action recognition in the EPIC-Kitchens, EGTEA Gaze+, ADL and Charades-EGO datasets demonstrate the improvements of our approach over single-dataset baselines. On EGTEA we surpass the current state-of-the-art by 2.47 percent. We further illustrate the cross-dataset task correlations that emerge automatically with our novel training scheme.

Deep learning approaches have been established as the main methodology for video classification and recognition. Recently, 3-dimensional convolutions have been used to achieve state-of-the-art performance in many challenging video datasets. Because of the high level of complexity of these methods, as the convolution operations are also extended to an additional dimension in order to extract features from it as well, providing a visualization for the signals that the network interpret as informative, is a challenging task. An effective notion of understanding the network's innerworkings would be to isolate the spatio-temporal regions on the video that the network finds most informative. We propose a method called Saliency Tubes which demonstrate the foremost points and regions in both frame level and over time that are found to be the main focus points of the network. We demonstrate our findings on widely used datasets for thirdperson and egocentric action classification and enhance the set of methods and visualizations that improve 3D Convolutional Neural Networks (CNNs) intelligibility. Our code1and a demo video2are also available.