We present FoveaBox, an accurate, flexible and completely anchor-free framework for object detection. While almost all state-of-the-art object detectors utilize the predefined anchors to enumerate possible locations, scales and aspect ratios for the search of the objects, their performance and generalization ability are also limited to the design of anchors. Instead, FoveaBox directly learns the object existing possibility and the bounding box coordinates without anchor reference. This is achieved by: (a) predicting category-sensitive semantic maps for the object existing possibility, and (b) producing category-agnostic bounding box for each position that potentially contains an object. The scales of target boxes are naturally associated with feature pyramid representations for each input image. Without bells and whistles, FoveaBox achieves state-of-the-art single model performance of 42.1 AP on the standard COCO detection benchmark. Specially for the objects with arbitrary aspect ratios, FoveaBox brings in significant improvement compared to the anchor-based detectors. More surprisingly, when it is challenged by the stretched testing images, FoveaBox shows great robustness and generalization ability to the changed distribution of bounding box shapes. The code will be made publicly available.
Recently, researchers proposed various low-precision gradient compression, for efficient communication in large-scale distributed optimization. Based on these work, we try to reduce the communication complexity from a new direction. We pursue an ideal bijective mapping between two spaces of gradient distribution, so that the mapped gradient carries greater information entropy after the compression. In our setting, all servers should share a reference gradient in advance, and they communicate via the normalized gradients, which are the subtraction or quotient, between current gradients and the reference. To obtain a reference vector that yields a stronger signal-to-noise ratio, dynamically in each iteration, we extract and fuse information from the past trajectory in hindsight, and search for an optimal reference for compression. We name this to be the trajectory-based normalized gradients (TNG). It bridges the research from different societies, like coding, optimization, systems, and learning. It is easy to implement and can universally combine with existing algorithms. Our experiments on benchmarking hard non-convex functions, convex problems like logistic regression demonstrate that TNG is more compression-efficient for communication of distributed optimization of general functions.
We present consistent optimization for single stage object detection. Previous works of single stage object detectors usually rely on the regular, dense sampled anchors to generate hypothesis for the optimization of the model. Through an examination of the behavior of the detector, we observe that the misalignment between the optimization target and inference configurations has hindered the performance improvement. We propose to bride this gap by consistent optimization, which is an extension of the traditional single stage detector's optimization strategy. Consistent optimization focuses on matching the training hypotheses and the inference quality by utilizing of the refined anchors during training. To evaluate its effectiveness, we conduct various design choices based on the state-of-the-art RetinaNet detector. We demonstrate it is the consistent optimization, not the architecture design, that yields the performance boosts. Consistent optimization is nearly cost-free, and achieves stable performance gains independent of the model capacities or input scales. Specifically, utilizing consistent optimization improves RetinaNet from 39.1 AP to 40.1 AP on COCO dataset without any bells or whistles, which surpasses the accuracy of all existing state-of-the-art one-stage detectors when adopting ResNet-101 as backbone. The code will be made available.
For recovering 3D object poses from 2D images, a prevalent method is to pre-train an over-complete dictionary $\mathcal D=\{B_i\}_i^D$ of 3D basis poses. During testing, the detected 2D pose $Y$ is matched to dictionary by $Y \approx \sum_i M_i B_i$ where $\{M_i\}_i^D=\{c_i \Pi R_i\}$, by estimating the rotation $R_i$, projection $\Pi$ and sparse combination coefficients $c \in \mathbb R_{+}^D$. In this paper, we propose non-convex regularization $H(c)$ to learn coefficients $c$, including novel leaky capped $\ell_1$-norm regularization (LCNR), \begin{align*} H(c)=\alpha \sum_{i } \min(|c_i|,\tau)+ \beta \sum_{i } \max(| c_i|,\tau), \end{align*} where $0\leq \beta \leq \alpha$ and $0<\tau$ is a certain threshold, so the invalid components smaller than $\tau$ are composed with larger regularization and other valid components with smaller regularization. We propose a multi-stage optimizer with convex relaxation and ADMM. We prove that the estimation error $\mathcal L(l)$ decays w.r.t. the stages $l$, \begin{align*} Pr\left(\mathcal L(l) < \rho^{l-1} \mathcal L(0) + \delta \right) \geq 1- \epsilon, \end{align*} where $0< \rho <1, 0<\delta, 0<\epsilon \ll 1$. Experiments on large 3D human datasets like H36M are conducted to support our improvement upon previous approaches. To the best of our knowledge, this is the first theoretical analysis in this line of research, to understand how the recovery error is affected by fundamental factors, e.g. dictionary size, observation noises, optimization times. We characterize the trade-off between speed and accuracy towards real-time inference in applications.
Despite huge success in the image domain, modern detection models such as Faster R-CNN have not been used nearly as much for video analysis. This is arguably due to the fact that detection models are designed to operate on single frames and as a result do not have a mechanism for learning motion representations directly from video. We propose a learning procedure that allows detection models such as Faster R-CNN to learn motion features directly from the RGB video data while being optimized with respect to a pose estimation task. Given a pair of video frames---Frame A and Frame B---we force our model to predict human pose in Frame A using the features from Frame B. We do so by leveraging deformable convolutions across space and time. Our network learns to spatially sample features from Frame B in order to maximize pose detection accuracy in Frame A. This naturally encourages our network to learn motion offsets encoding the spatial correspondences between the two frames. We refer to these motion offsets as DiMoFs (Discriminative Motion Features). In our experiments we show that our training scheme helps learn effective motion cues, which can be used to estimate and localize salient human motion. Furthermore, we demonstrate that as a byproduct, our model also learns features that lead to improved pose detection in still-images, and better keypoint tracking. Finally, we show how to leverage our learned model for the tasks of spatiotemporal action localization and fine-grained action recognition.
We propose a light-weight video frame interpolation algorithm. Our key innovation is an instance-level supervision that allows information to be learned from the high-resolution version of similar objects. Our experiment shows that the proposed method can generate state-of-art results across different datasets, with fractional computation resources (time and memory) with competing methods. Given two image frames, a cascade network creates an intermediate frame with 1) a flow-warping module that computes large bi-directional optical flow and creates an interpolated image via flow-based warping, followed by 2) an image synthesis module to make fine-scale corrections. In the learning stage, object detection proposals are generated on the interpolated image. Lower resolution objects are zoomed into, and the learning algorithms using an adversarial loss trained on high-resolution objects to guide the system towards the instance-level refinement corrects details of object shape and boundaries. As all our proposed network modules are fully convolutional, our proposed system can be trained end-to-end.
We propose a Spatiotemporal Sampling Network (STSN) that uses deformable convolutions across time for object detection in videos. Our STSN performs object detection in a video frame by learning to spatially sample features from the adjacent frames. This naturally renders the approach robust to occlusion or motion blur in individual frames. Our framework does not require additional supervision, as it optimizes sampling locations directly with respect to object detection performance. Our STSN outperforms the state-of-the-art on the ImageNet VID dataset and compared to prior video object detection methods it uses a simpler design, and does not require optical flow data for training.
The per-pixel cross-entropy loss (CEL) has been widely used in structured output prediction tasks as a spatial extension of generic image classification. However, its i.i.d. assumption neglects the structural regularity present in natural images. Various attempts have been made to incorporate structural reasoning mostly through structure priors in a cooperative way where co-occuring patterns are encouraged. We, on the other hand, approach this problem from an opposing angle and propose a new framework for training such structured prediction networks via an adversarial process, in which we train a structure analyzer that provides the supervisory signals, the adversarial structure matching loss (ASML). The structure analyzer is trained to maximize ASML, or to exaggerate recurring structural mistakes usually among co-occurring patterns. On the contrary, the structured output prediction network is trained to reduce those mistakes and is thus enabled to distinguish fine-grained structures. As a result, training structured output prediction networks using ASML reduces contextual confusion among objects and improves boundary localization. We demonstrate that ASML outperforms its counterpart CEL especially in context and boundary aspects on figure-ground segmentation and semantic segmentation tasks with various base architectures, such as FCN, U-Net, DeepLab, and PSPNet.
We present a model that uses a single first-person image to generate an egocentric basketball motion sequence in the form of a 12D camera configuration trajectory, which encodes a player's 3D location and 3D head orientation throughout the sequence. To do this, we first introduce a future convolutional neural network (CNN) that predicts an initial sequence of 12D camera configurations, aiming to capture how real players move during a one-on-one basketball game. We also introduce a goal verifier network, which is trained to verify that a given camera configuration is consistent with the final goals of real one-on-one basketball players. Next, we propose an inverse synthesis procedure to synthesize a refined sequence of 12D camera configurations that (1) sufficiently matches the initial configurations predicted by the future CNN, while (2) maximizing the output of the goal verifier network. Finally, by following the trajectory resulting from the refined camera configuration sequence, we obtain the complete 12D motion sequence. Our model generates realistic basketball motion sequences that capture the goals of real players, outperforming standard deep learning approaches such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and generative adversarial networks (GANs).
We present a first-person method for cooperative basketball intention prediction: we predict with whom the camera wearer will cooperate in the near future from unlabeled first-person images. This is a challenging task that requires inferring the camera wearer's visual attention, and decoding the social cues of other players. Our key observation is that a first-person view provides strong cues to infer the camera wearer's momentary visual attention, and his/her intentions. We exploit this observation by proposing a new cross-model EgoSupervision learning scheme that allows us to predict with whom the camera wearer will cooperate in the near future, without using manually labeled intention labels. Our cross-model EgoSupervision operates by transforming the outputs of a pretrained pose-estimation network, into pseudo ground truth labels, which are then used as a supervisory signal to train a new network for a cooperative intention task. We evaluate our method, and show that it achieves similar or even better accuracy than the fully supervised methods do.