RetinaViT: Efficient Visual Backbone for Online Video Streams

In online video understanding, which has a wide range of real-world applications, inference speed is crucial. Many approaches involve frame-level visual feature extraction, which often represents the biggest bottleneck. We propose RetinaViT, an efficient method for extracting frame-level visual features in an online video stream, aiming to fundamentally enhance the efficiency of online video understanding tasks. RetinaViT is composed of efficiently approximated Transformer blocks that only take changed tokens (event tokens) as queries and reuse the already processed tokens from the previous timestep for the others. Furthermore, we restrict keys and values to the spatial neighborhoods of event tokens to further improve efficiency. RetinaViT involves tuning multiple parameters, which we determine through a multi-step process. During model training, we randomly vary these parameters and then perform black-box optimization to maximize accuracy and efficiency on the pre-trained model. We conducted extensive experiments on various online video recognition tasks, including action recognition, pose estimation, and object segmentation, validating the effectiveness of each component in RetinaViT and demonstrating improvements in the speed/accuracy trade-off compared to baselines. In particular, for action recognition, RetinaViT built on ViT-B16 reduces inference time by approximately 61.9% on the CPU and 50.8% on the GPU, while achieving slight accuracy improvements rather than degradation.