MARS - Naver Labs Europe
loader image

MARS: Motion-Augmented RGB Stream for Action Recognition


Mars cartwheel gif

The test code and models are released under the MIT license.

Both the code and models are on github.

Publication and blog

The CVPR 2019 publication related to the codes and models is MARS: Motion-Augmented RGB Stream for Action Recognition [PDF]

Authors: Nieves Crasto1, Philippe Weinzaepfel1, Karteek Alahari2, Cordelia Schmid2 [1NAVER LABS Europe  2Inria]

Look at it differently – there’s also a quick read blog article on MARS

Information related to code and models

Most state-of-the-art methods for action recognition consist of a two-stream architecture with 3D convolutions: an appearance stream for RGB frames and a motion stream for  optical flow frames. Although combining flow with RGB improves the performance, the cost of computing accurate optical flow is high, and increases action recognition latency. This limits the usage of two-stream approaches in real-world applications requiring low latency. In this paper, we  introduce two learning approaches to train a standard 3D CNN, operating on RGB frames, that mimics the motion stream, and as a result avoids flow computation at test time. First, by minimizing a feature-based loss compared to the Flow stream, we show that the network reproduces the motion stream with high fidelity. Second, to leverage both appearance and motion information effectively, we train with a linear combination of the feature-based loss and the standard cross-entropy loss for action recognition. We denote the stream trained using this combined loss as Motion-Augmented  RGB Stream (MARS). As a single stream, MARS performs better than RGB or Flow alone, for instance with 72.7% accuracy on Kinetics compared to 72.0% and 65.6% with RGB and Flow streams respectively.



 title={{MARS: Motion-Augmented RGB Stream for Action Recognition}},
 author={Crasto, Nieves and Weinzaepfel, Philippe and Alahari, Karteek and Schmid, Cordelia},