We present an online approach to efficiently and simultaneously detect and track 2D poses of multiple people in a video sequence. We build upon Part Affinity Field (PAF) representation designed for static images, and propose an architecture that can encode and predict Spatio-Temporal Affinity Fields (STAF) across a video sequence. In particular, we propose a novel temporal topology cross-linked across limbs which can consistently handle body motions of a wide range of magnitudes. Additionally, we make the overall approach recurrent in nature, where the network ingests STAF heatmaps from previous frames and estimates those for the current frame. Our approach uses only online inference and tracking, and is currently the fastest and the most accurate bottom-up approach that is runtime-invariant to the number of people in the scene and accuracy-invariant to input frame rate of camera. Running at ~30 fps on a single GPU at single scale, it achieves highly competitive results on the PoseTrack benchmarks.

We present the first single-network approach for 2D~whole-body pose estimation, which entails simultaneous localization of body, face, hands, and feet keypoints. Due to the bottom-up formulation, our method maintains constant real-time performance regardless of the number of people in the image. The network is trained in a single stage using multi-task learning, through an improved architecture which can handle scale differences between body/foot and face/hand keypoints. Our approach considerably improves upon OpenPose~\cite{cao2018openpose}, the only work so far capable of whole-body pose estimation, both in terms of speed and global accuracy. Unlike OpenPose, our method does not need to run an additional network for each hand and face candidate, making it substantially faster for multi-person scenarios. This work directly results in a reduction of computational complexity for applications that require 2D whole-body information (e.g., VR/AR, re-targeting). In addition, it yields higher accuracy, especially for occluded, blurry, and low resolution faces and hands. For code, trained models, and validation benchmarks, visit our project page: https://github.com/CMU-Perceptual-Computing-Lab/openpose_train.

This paper proposes a person-centric and online approach to the challenging problem of localization and prediction of actions and interactions in videos. Typically, localization or recognition is performed in an offline manner where all the frames in the video are processed together. This prevents timely localization and prediction of actions and interactions - an important consideration for many tasks including surveillance and human-machine interaction. In our approach, we estimate human poses at each frame and train discriminative appearance models using the superpixels inside the pose bounding boxes. Since the pose estimation per frame is inherently noisy, the conditional probability of pose hypotheses at current time-step (frame) is computed using pose estimations in the current frame and their consistency with poses in the previous frames. Next, both the superpixel and pose-based foreground likelihoods are used to infer the location of actors at each time through a Conditional Random Field enforcing spatio-temporal smoothness in color, optical flow, motion boundaries and edges among superpixels. The issue of visual drift is handled by updating the appearance models, and refining poses using motion smoothness on joint locations, in an online manner. For online prediction of action/interaction confidences, we propose an approach based on Structural SVM that operates on short video segments, and is trained with the objective that confidence of an action or interaction increases as time passes in a positive training clip. Lastly, we quantify the performance of both detection and prediction together, and analyze how the prediction accuracy varies as a time function of observed action/interaction at different levels of detection performance. Our experiments on several datasets suggest that despite using only a few frames to localize actions/interactions at each time instant, we are able to obtain competitive results to state-of-the-art offline methods.

This paper presents a new approach for the challenging problem of geo-localization using image matching in a structured database of city-wide reference images with known GPS coordinates. We cast the geo-localization as a clustering problem of local image features. Akin to existing approaches to the problem, our framework builds on low-level features which allow local matching between images. For each local feature in the query image, we find its approximate nearest neighbors in the reference set. Next, we cluster the features from reference images using Dominant Set clustering, which affords several advantages over existing approaches. First, it permits variable number of nodes in the cluster, which we use to dynamically select the number of nearest neighbors for each query feature based on its discrimination value. Second, this approach is several orders of magnitude faster than existing approaches. Thus, we obtain multiple clusters (different local maximizers) and obtain a robust final solution to the problem using multiple weak solutions through constrained Dominant Set clustering on global image features, where we enforce the constraint that the query image must be included in the cluster. This second level of clustering also bypasses heuristic approaches to voting and selecting the reference image that matches to the query. We evaluate the proposed framework on an existing dataset of 102k street view images as well as a new larger dataset of 300k images, and show that it outperforms the state-of-the-art by 20 and 7 percent, respectively, on the two datasets.

This paper proposes a novel approach to tackle the challenging problem of 'online action localization' which entails predicting actions and their locations as they happen in a video. Typically, action localization or recognition is performed in an offline manner where all the frames in the video are processed together and action labels are not predicted for the future. This disallows timely localization of actions - an important consideration for surveillance tasks. In our approach, given a batch of frames from the immediate past in a video, we estimate pose and oversegment the current frame into superpixels. Next, we discriminatively train an actor foreground model on the superpixels using the pose bounding boxes. A Conditional Random Field with superpixels as nodes, and edges connecting spatio-temporal neighbors is used to obtain action segments. The action confidence is predicted using dynamic programming on SVM scores obtained on short segments of the video, thereby capturing sequential information of the actions. The issue of visual drift is handled by updating the appearance model and pose refinement in an online manner. Lastly, we introduce a new measure to quantify the performance of action prediction (i.e. online action localization), which analyzes how the prediction accuracy varies as a function of observed portion of the video. Our experiments suggest that despite using only a few frames to localize actions at each time instant, we are able to predict the action and obtain competitive results to state-of-the-art offline methods.

This paper presents an efficient approach for localizing actions by learning contextual relations, in the form of relative locations between different video regions. We begin by over-segmenting the videos into supervoxels, which have the ability to preserve action boundaries and also reduce the complexity of the problem. Context relations are learned during training which capture displacements from all the supervoxels in a video to those belonging to foreground actions. Then, given a testing video, we select a supervoxel randomly and use the context information acquired during training to estimate the probability of each supervoxel belonging to the foreground action. The walk proceeds to a new supervoxel and the process is repeated for a few steps. This "context walk" generates a conditional distribution of an action over all the supervoxels. A Conditional Random Field is then used to find action proposals in the video, whose confidences are obtained using SVMs. We validated the proposed approach on several datasets and show that context in the form of relative displacements between supervoxels can be extremely useful for action localization. This also results in significantly fewer evaluations of the classifier, in sharp contrast to the alternate sliding window approaches.

Human detection in dense crowds is an important problem, as it is a prerequisite to many other visual tasks, such as tracking, counting, action recognition or anomaly detection in behaviors exhibited by individuals in a dense crowd. This problem is challenging due to the large number of individuals, small apparent size, severe occlusions and perspective distortion. However, crowded scenes also offer contextual constraints that can be used to tackle these challenges. In this paper, we explore context for human detection in dense crowds in the form of a locally-consistent scale prior which captures the similarity in scale in local neighborhoods and its smooth variation over the image. Using the scale and confidence of detections obtained from an underlying human detector, we infer scale and confidence priors using Markov Random Field. In an iterative mechanism, the confidences of detection hypotheses are modified to reflect consistency with the inferred priors, and the priors are updated based on the new detections. The final set of detections obtained are then reasoned for occlusion using Binary Integer Programming where overlaps and relations between parts of individuals are encoded as linear constraints. Both human detection and occlusion reasoning in proposed approach are solved with local neighbor-dependent constraints, thereby respecting the inter-dependence between individuals characteristic to dense crowd analysis. In addition, we propose a mechanism to detect different combinations of body parts without requiring annotations for individual combinations. We performed experiments on a new and extremely challenging dataset of dense crowd images showing marked improvement over the underlying human detector.

A video captures a sequence and interactions of concepts that can be static, for instance, objects or scenes, or dynamic, such as actions. For large datasets containing hundreds of thousands of images or videos, it is impractical to manually annotate all the concepts, or all the instances of a single concept. However, a dictionary with visually-distinct elements can be created automatically from unlabeled videos which can capture and express the entire dataset. The downside to this machine-discovered dictionary is meaninglessness, i.e., its elements are devoid of semantics and interpretation. In this paper, we present an approach that leverages the strengths of semantic concepts and the machine-discovered DOVE by learning a relationship between them. Since instances of a semantic concept share visual similarity, the proposed approach uses soft-consensus regularization to learn the mapping that enforces instances from each semantic concept to have similar representations. The testing is performed by projecting the query onto the DOVE as well as new representations of semantic concepts from training, with non-negativity and unit summation constraints for probabilistic interpretation. We tested our formulation on TRECVID MED and SIN tasks, and obtained encouraging results.

The real world image databases such as Flickr are characterized by continuous addition of new images. The recent approaches for image annotation, i.e. the problem of assigning tags to images, have two major drawbacks. First, either models are learned using the entire training data, or to handle the issue of dataset imbalance, tag-specific discriminative models are trained. Such models become obsolete and require relearning when new images and tags are added to database. Second, the task of feature-fusion is typically dealt using ad-hoc approaches. In this paper, we present a weighted extension of Multi-view Non-negative Matrix Factorization (NMF) to address the aforementioned drawbacks. The key idea is to learn query-specific generative model on the features of nearest-neighbors and tags using the proposed NMF-KNN approach which imposes consensus constraint on the coefficient matrices across different features. This results in coefficient vectors across features to be consistent and, thus, naturally solves the problem of feature fusion, while the weight matrices introduced in the proposed formulation alleviate the issue of dataset imbalance. Furthermore, our approach, being query-specific, is unaffected by addition of images and tags in a database. We tested our method on two datasets used for evaluation of image annotation and obtained competitive results.

We propose to leverage multiple sources of information to compute an estimate of the number of individuals present in an extremely dense crowd visible in a single image. Due to problems including perspective, occlusion, clutter, and few pixels per person, counting by human detection in such images is almost impossible. Instead, our approach relies on multiple sources such as low confidence head detections, repetition of texture elements (using SIFT), and frequency-domain analysis to estimate counts, along with confidence associated with observing individuals, in an image region. Secondly, we employ a global consistency constraint on counts using Markov Random Field. This caters for disparity in counts in local neighborhoods and across scales. We tested our approach on a new dataset of fifty crowd images containing 64K annotated humans, with the head counts ranging from 94 to 4543. This is in stark contrast to datasets used for existing methods which contain not more than tens of individuals. We experimentally demonstrate the efficacy and reliability of the proposed approach by quantifying the counting performance.