Modality Distillation with Multiple Stream Networks for Action Recognition

Abstract

Diverse input data modalities can provide complementary cues for several tasks, usually leading to more robust algorithms and better performance. However, while a (training) dataset could be accurately designed to include a variety of sensory inputs, it is often the case that not all modalities could be available in real life (testing) scenarios, where a model has to be deployed. This raises the challenge of how to learn robust representations leveraging multimodal data in the training stage, while considering limitations at test time, such as noisy or missing modalities. This paper presents a new approach for multimodal video action recognition, developed within the unified frameworks of distillation and privileged information, named generalized distillation. Particularly, we consider the case of learning representations from depth and RGB videos, while relying on RGB data only at test time. We propose a new approach to train an hallucination network that learns to distill depth features through multiplicative connections of spatiotemporal representations, leveraging soft labels and hard labels, as well as distance between feature maps. We report state-of-the-art results on video action classification on the largest multimodal dataset available for this task, the NTU RGB+D. Code available at https://github.com/ncgarcia/ modality-distillation

Learning with Privileged Information via Adversarial Discriminative Modality Distillation

“…Inspired by the generalized distillation paradigm, we follow a staged learning procedure, where the "teacher" net is trained first (Step 1) and separately from the "student" (Step 2). This is in contrast with [10], where everything is learned end-to-end, but in line with [11], where separated learning steps proved to be more effective.…”

“…In this context, the closest works to our approach are [10] by Hoffman et al and [11] by Garcia et al…”

“…Moreover, the authors encouraged the learning by design, by using multiplier cross-stream connections [19], and extended the method to video action recognition. Differently from [11] and [10], we propose an adversarial strategy to learn the hallucination stream, which alleviates the need for balancing losses or tuning hyperparameters.…”

“…• We propose a new approach to learn a hallucination network within a multimodal-stream network architecture: it consists in an adversarial learning strategy that exploits multiple data modalities at training while using only one at test time. It proved to outperform its distance-based method counterparts [10], [11], and to augment its flexibility by being agnostic to components like distance metrics, data modalities, and size of the hallucinated feature vectors.…”

Cross-View Policy Learning for Street Navigation

“…The hallucination network in [18] was trained on an existing modality to regress the missing modality using L 2 loss, and leveraged multiple such losses for multiple tasks. This work has been extended by Garcia et al by adding L 2 losses for reconstructing all layers of the depth network and a cross entropy distillation loss for a missing network [12]. Finally, Luo et al [26] learned the direction of distillation between modalities, considering a cosine distillation loss and a representation loss.…”