Building a Size Constrained Predictive Model for Video Classification

Miha Skalic, David Austin; Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018, pp. 0-0

Herein we present the solution to the 2nd YouTube-8M video understanding challenge which placed 1st. Competition participants were tasked with building a size constrained video labeling model with a model size of less than 1GB. Our final solution consists of several submodels belonging to Fisher vectors, NetVlad, Deep Bag of Frames and Recurrent neural networks model families. To make the classifier efficient under size constraints we introduced model distillation, partial weights quantization and training with exponential moving average.