The most performant spatio-temporal action localisation models use external person proposals and complex external memory banks. We propose a fully end-to-end, purely-transformer based model that directly ingests an input video, and outputs tubelets - a sequence of bounding boxes and the action classes at each frame. Our flexible model can be trained with either sparse bounding-box supervision on individual frames, or full tubelet annotations. And in both cases, it predicts coherent tubelets as the output. Moreover, our end-to-end model requires no additional pre-processing in the form of proposals, or post-processing in terms of non-maximal suppression. We perform extensive ablation experiments, and significantly advance the state-of-the-art on five different spatio-temporal action localisation benchmarks with both sparse keyframes and full tubelet annotations.

An ideal model for dense video captioning - predicting captions localized temporally in a video - should be able to handle long input videos, predict rich, detailed textual descriptions, and be able to produce outputs before processing the entire video. Current state-of-the-art models, however, process a fixed number of downsampled frames, and make a single full prediction after seeing the whole video. We propose a streaming dense video captioning model that consists of two novel components: First, we propose a new memory module, based on clustering incoming tokens, which can handle arbitrarily long videos as the memory is of a fixed size. Second, we develop a streaming decoding algorithm that enables our model to make predictions before the entire video has been processed. Our model achieves this streaming ability, and significantly improves the state-of-the-art on three dense video captioning benchmarks: ActivityNet, YouCook2 and ViTT. Our code is released at https://github.com/google-research/scenic.

While large-scale image-text pretrained models such as CLIP have been used for multiple video-level tasks on trimmed videos, their use for temporal localization in untrimmed videos is still a relatively unexplored task. We design a new approach for this called UnLoc, which uses pretrained image and text towers, and feeds tokens to a video-text fusion model. The output of the fusion module are then used to construct a feature pyramid in which each level connects to a head to predict a per-frame relevancy score and start/end time displacements. Unlike previous works, our architecture enables Moment Retrieval, Temporal Localization, and Action Segmentation with a single stage model, without the need for action proposals, motion based pretrained features or representation masking. Unlike specialized models, we achieve state of the art results on all three different localization tasks with a unified approach. Code is available at: https://github.com/google-research/scenic.

Learning visual knowledge from massive weakly-labeled web videos has attracted growing research interests thanks to the large corpus of easily accessible video data on the Internet. However, for video action recognition, the action of interest might only exist in arbitrary clips of untrimmed web videos, resulting in high label noises in the temporal space. To address this challenge, we introduce a new method for pre-training video action recognition models using queried web videos. Instead of trying to filter out potential noises, we propose to provide fine-grained supervision signals by defining the concept of Sub-Pseudo Label (SPL). Specifically, SPL spans out a new set of meaningful "middle ground" label space constructed by extrapolating the original weak labels during video querying and the prior knowledge distilled from a teacher model. Consequently, SPL provides enriched supervision for video models to learn better representations and improves data utilization efficiency of untrimmed videos. We validate the effectiveness of our method on four video action recognition datasets and a weakly-labeled image dataset. Experiments show that SPL outperforms several existing pre-training strategies and the learned representations lead to competitive results on several benchmarks.

Video understanding requires reasoning at multiple spatiotemporal resolutions – from short fine-grained motions to events taking place over longer durations. Although transformer architectures have recently advanced the state-of-the-art, they have not explicitly modelled different spatiotemporal resolutions. To this end, we present Multiview Transformers for Video Recognition (MTV). Our model consists of separate encoders to represent different views of the input video with lateral connections to fuse information across views. We present thorough ablation studies of our model and show that MTV consistently performs better than single-view counterparts in terms of accuracy and computational cost across a range of model sizes. Furthermore, we achieve state-of-the-art results on six standard datasets, and improve even further with large-scale pretraining. Code and checkpoints are available at: https://github.com/google-research/scenic.

In recent years, many works in the video action recognition literature have shown that two stream models (combining spatial and temporal input streams) are necessary for achieving state-of-the-art performance. In this paper we show the benefits of including yet another stream based on human pose estimated from each frame — specifically by rendering pose on input RGB frames. At first blush, this additional stream may seem redundant given that human pose is fully determined by RGB pixel values — however we show (perhaps surprisingly) that this simple and flexible addition can provide complementary gains. Using this insight, we propose a new model, which we dub PERF-Net (short for Pose Empowered RGB-Flow Net), which combines this new pose stream with the standard RGB and flow based input streams via distillation techniques and show that our model outperforms the state-of-the-art by a large margin in a number of human action recognition datasets while not requiring flow or pose to be explicitly computed at inference time. The proposed pose stream is also part of the winner solution of the ActivityNet Kinetics Challenge 2020 [1].

Video object segmentation targets segmenting a specific object throughout a video sequence when given only an annotated first frame. Recent deep learning based approaches find it effective to fine-tune a general-purpose segmentation model on the annotated frame using hundreds of iterations of gradient descent. Despite the high accuracy that these methods achieve, the fine-tuning process is inefficient and fails to meet the requirements of real world applications. We propose a novel approach that uses a single forward pass to adapt the segmentation model to the appearance of a specific object. Specifically, a second meta neural network named modulator is trained to manipulate the intermediate layers of the segmentation network given limited visual and spatial information of the target object. The experiments show that our approach is 70× faster than fine-tuning approaches and achieves similar accuracy. Our model and code have been released at https://github.com/linjieyangsc/video_seg.

Mathematical optimization plays a fundamental role in solving many problems in computer vision (e.g., camera calibration, image alignment, structure from motion). It is generally accepted that second order descent methods are the most robust, fast, and reliable approaches for nonlinear optimization of a general smooth function. However, in the context of computer vision, second order descent methods have two main drawbacks: 1) the function might not be analytically differentiable and numerical approximations are impractical, and 2) the Hessian may be large and not positive definite. Recently, Supervised Descent Method (SDM), a method that learns the “weighted averaged gradients” in a supervised manner has been proposed to solve these issues. However, SDM is a local algorithm and it is likely to average conflicting gradient directions. This paper proposes Global SDM (GSDM), an extension of SDM that divides the search space into regions of similar gradient directions. GSDM provides a better and more efficient strategy to minimize non-linear least squares functions in computer vision problems. We illustrate the effectiveness of GSDM in two problems: non-rigid image alignment and extrinsic camera calibration.

Distracted driving is one of the main causes of vehicle collisions in the United States. Passively monitoring a driver's activities constitutes the basis of an automobile safety system that can potentially reduce the number of accidents by estimating the driver's focus of attention. This paper proposes an inexpensive vision-based system to accurately detect Eyes Off the Road (EOR). The system has three main components: 1) robust facial feature tracking; 2) head pose and gaze estimation; and 3) 3-D geometric reasoning to detect EOR. From the video stream of a camera installed on the steering wheel column, our system tracks facial features from the driver's face. Using the tracked landmarks and a 3-D face model, the system computes head pose and gaze direction. The head pose estimation algorithm is robust to nonrigid face deformations due to changes in expressions. Finally, using a 3-D geometric analysis, the system reliably detects EOR.

Many computer vision problems (e.g., camera calibration, image alignment, structure from motion) are solved through a nonlinear optimization method. It is generally accepted that 2nd order descent methods are the most robust, fast and reliable approaches for nonlinear optimization of a general smooth function. However, in the context of computer vision, 2nd order descent methods have two main drawbacks: (1) The function might not be analytically differentiable and numerical approximations are impractical. (2) The Hessian might be large and not positive definite. To address these issues, this paper proposes a Supervised Descent Method (SDM) for minimizing a Non-linear Least Squares (NLS) function. During training, the SDM learns a sequence of descent directions that minimizes the mean of NLS functions sampled at different points. In testing, SDM minimizes the NLS objective using the learned descent directions without computing the Jacobian nor the Hessian. We illustrate the benefits of our approach in synthetic and real examples, and show how SDM achieves state-of-the-art performance in the problem of facial feature detection. The code is available at www.humansensing.cs. cmu.edu/intraface.

We address the problem of understanding scenes from 3-D laser scans via per-point assignment of semantic labels. In order to mitigate the difficulties of using a graphical model for modeling the contextual relationships among the 3-D points, we instead propose a multi-stage inference procedure to capture these relationships. More specifically, we train this procedure to use point cloud statistics and learn relational information (e.g., tree-trunks are below vegetation) over fine (point-wise) and coarse (region-wise) scales. We evaluate our approach on three different datasets, that were obtained from different sensors, and demonstrate improved performance.

Semantic 3D models of buildings encode the geometry as well as the identity of key components of a facility, such as walls, floors, and ceilings. Manually constructing such a model is a time-consuming and error-prone process. Our goal is to automate this process using 3D point data from a laser scanner. Our hypothesis is that contextual information is important to reliable performance in unmodified environments, which are often highly cluttered. We use a Conditional Random Field (CRF) model to discover and exploit contextual information, classifying planar patches extracted from the point cloud data. We compare the results of our context-based CRF algorithm with a context-free method based on L2 norm regularized Logistic Regression (RLR). We find that using certain contextual information along with local features leads to better classification results. 1