Vision Transformer (ViT) has demonstrated promising performance in various computer vision tasks, and recently attracted a lot of research attention. Many recent works have focused on proposing new architectures to improve ViT and deploying it into real-world applications. However, little effort has been made to analyze and understand ViT’s architecture design space and its implication for hardware costs on different devices. In this work, by simply scaling ViT’s depth, width, input size, and other basic configurations, we show that a scaled vanilla ViT model without bells and whistles can achieve comparable or superior accuracy-efficiency trade-off than most of the latest ViT variants. Specifically, compared with DeiT-Tiny, our scaled model achieves a ↑ 1.9% higher ImageNet top-1 accuracy under the same FLOPs and a ↑ 3.7% better ImageNet top-1 accuracy under the same latency on an NVIDIA Edge GPU TX2. Motivated by this, we further investigate the extracted scaling strategies from the following two aspects: (1) can these scaling strategies be transferred across different real hardware devices ? and (2) can these scaling strategies be transferred to different ViT variants and tasks ?. For (1), our exploration, based on various devices with different resource budgets, indicates that the transferability effectiveness depends on the underlying device together with its corresponding deployment tool. For (2), we validate the effective transferability of the aforementioned scaling strategies obtained from a vanilla ViT model on top of an image classification task to the PiT model, a strong ViT variant targeting efficiency as well as object detection and video classification tasks. In particular, when transferred to PiT, our scaling strategies lead to a boosted ImageNet top-1 accuracy of from 74.6% to 76.7% (↑ 2.1%) under the same 0.7G FLOPs. When transferred to the COCO object detection task, the average precision is boosted by ↑ 0.7% under a similar throughput on a V100 GPU.

Temporal action localization has been one of the most popular tasks in video understanding, due to the importance of detecting action instances in videos. However, not much progress has been made on extending it to work in an on-line fashion, although many video related tasks can benefit by going online with the growing video streaming services. To this end, we introduce a new task called Online Temporal Action Localization (On-TAL), in which the goal is to immediately detect action instances from an untrimmed streaming video. The online setting makes the new task very challenging as the actionness decision for every frame has to be made without access to future frames and also because post-processing methods cannot be used to modify past action proposals. We propose a novel framework, Context-Aware Actionness Grouping (CAG) as a solution for On-TAL and train it with the imitation learning algorithm, which allows us to avoid sophisticated reward engineering. Evaluation of our work on THUMOS14 and Activitynet1.3 shows significant improvement over non-naive baselines, demonstrating the effectiveness of our approach. As a by-product, our method can also be used for the Online Detection of Action Start (ODAS), in which our method also outperforms previous state-of-the-art models.