CVPR 2020 Papers

Neural architecture search (NAS) approaches aim at automatically finding novel CNN architectures that fit computational constraints while maintaining a good performance on the target platform. We introduce a novel efficient one-shot NAS approach to optimally search for channel numbers, given latency constraints on a specific hardware. We first show that we can use a black-box approach to estimate a realistic latency model for a specific inference platform, without the need for low-level access to the inference computation. Then, we design a pairwise MRF to score any channel configuration and use dynamic programming to efficiently decode the best performing configuration, yielding an optimal solution for the network width search. Finally, we propose an adaptive channel configuration sampling scheme to gradually specialize the training phase to the target computational constraints. Experiments on ImageNet classification show that our approach can find networks fitting the resource constraints on different target platforms while improving accuracy over the state-of-the-art efficient networks.

Scene text detection and recognition has received increasing research attention. Existing methods can be roughly categorized into two groups: character-based and segmentation-based. These methods either are costly for character annotation or need to maintain a complex pipeline, which is often not suitable for real-time applications. Here we address the problem by proposing the Adaptive Bezier-Curve Network (\BeCan). Our contributions are three-fold: 1) For the first time, we adaptively fit oriented or curved text by a parameterized Bezier curve. 2) We design a novel BezierAlign layer for extracting accurate convolution features of a text instance with arbitrary shapes, significantly improving the precision compared with previous methods. 3) Compared with standard bounding box detection, our Bezier curve detection introduces negligible computation overhead, resulting in superiority of our method in both efficiency and accuracy. Experiments on oriented or curved benchmark datasets, namely Total-Text and CTW1500, demonstrate that \BeCan achieves state-of-the-art accuracy, meanwhile significantly improving the speed. In particular, on Total-Text, our real-time version is over 10 times faster than recent state-of-the-art methods with a competitive recognition accuracy. Code is available at https://git.io/AdelaiDet.

In this work, we develop a shared multi-attention model for multi-label zero-shot learning. We argue that designing attention mechanism for recognizing multiple seen and unseen labels in an image is a non-trivial task as there is no training signal to localize unseen labels and an image only contains a few present labels that need attentions out of thousands of possible labels. Therefore, instead of generating attentions for unseen labels which have unknown behaviors and could focus on irrelevant regions due to the lack of any training sample, we let the unseen labels select among a set of shared attentions which are trained to be label-agnostic and to focus on only relevant/foreground regions through our novel loss. Finally, we learn a compatibility function to distinguish labels based on the selected attention. We further propose a novel loss function that consists of three components guiding the attention to focus on diverse and relevant image regions while utilizing all attention features. By extensive experiments, we show that our method improves the state of the art by 2.9% and 1.4% F1 score on the NUS-WIDE and the large scale Open Images datasets, respectively.

Adversarial examples can cause catastrophic mistakes in Deep Neural Network (DNNs) based vision systems e.g., for classification, segmentation and object detection. The vulnerability of DNNs against such attacks can prove a major roadblock towards their real-world deployment. Transferability of adversarial examples demand generalizable defenses that can provide cross-task protection. Adversarial training that enhances robustness by modifying target model's parameters lacks such generalizability. On the other hand, different input processing based defenses fall short in the face of continuously evolving attacks. In this paper, we take the first step to combine the benefits of both approaches and propose a self-supervised adversarial training mechanism in the input space. By design, our defense is a generalizable approach and provides significant robustness against the unseen adversarial attacks (e.g. by reducing the success rate of translation-invariant ensemble attack from 82.6% to 31.9% in comparison to previous state-of-the-art). It can be deployed as a plug-and-play solution to protect a variety of vision systems, as we demonstrate for the case of classification, segmentation and detection.

Even as deep neural networks have become very effective for tasks in vision and perception, it remains difficult to explain and debug their behavior. In this paper, we present a programmatic and semantic approach to explaining, understanding, and debugging the correct and incorrect behaviors of a neural network based perception system. Our approach is semantic in that it employs a high-level representation of the distribution of environment scenarios that the detector is intended to work on. It is programmatic in that the representation is a program in a domain-specific probabilistic programming language using which synthetic data can be generated to train and test the neural network. We present a framework that assesses the performance of the neural network to identify correct and incorrect detections, extracts rules from those results that semantically characterizes the correct and incorrect scenarios, and then specializes the probabilistic program with those rules in order to more precisely characterize the scenarios in which the neural network operates correctly or not, without human intervention. We demonstrate our results using the Scenic probabilistic programming language and a neural network-based object detector. Our experiments show that it is possible to automatically generate compact rules that significantly increase the correct detection rate (or conversely the incorrect detection rate) of the network and can thus help with debugging and understanding its behavior.

Lacking rich and realistic data, learned single image denoising algorithms generalize poorly in real raw images that not resemble the data used for training. Although the problem can be alleviated by the heteroscedastic Gaussian noise model, the noise sources caused by digital camera electronics are still largely overlooked, despite their significant effect on raw measurement, especially under extremely low-light condition. To address this issue, we present a highly accurate noise formation model based on the characteristics of CMOS photosensors, thereby enabling us to synthesize realistic samples that better match the physics of image formation process. Given the proposed noise model, we additionally propose a method to calibrate the noise parameters for available modern digital cameras, which is simple and reproducible for any new device. We systematically study the generalizability of a neural network trained with existing schemes, by introducing a new low-light denoising dataset that covers many modern digital cameras from diverse brands. Extensive empirical results collectively show that by utilizing our proposed noise formation model, a network can reach the capability as if it had been trained with rich real data, which demonstrates the effectiveness of our noise formation model.

Training competitive deep video models is an order of magnitude slower than training their counterpart image models. Slow training causes long research cycles, which hinders progress in video understanding research. Following standard practice for training image models, video model training has used a fixed mini-batch shape: a specific number of clips, frames, and spatial size. However, what is the optimal shape? High resolution models perform well, but train slowly. Low resolution models train faster, but are less accurate. Inspired by multigrid methods in numerical optimization, we propose to use variable mini-batch shapes with different spatial-temporal resolutions that are varied according to a schedule. The different shapes arise from resampling the training data on multiple sampling grids. Training is accelerated by scaling up the mini-batch size and learning rate when shrinking the other dimensions. We empirically demonstrate a general and robust grid schedule that yields a significant out-of-the-box training speedup without a loss in accuracy for different models (I3D, non-local, SlowFast), datasets (Kinetics, Something-Something, Charades), and training settings (with and without pre-training, 128 GPUs or 1 GPU). As an illustrative example, the proposed multigrid method trains a ResNet-50 SlowFast network 4.5x faster (wall-clock time, same hardware) while also improving accuracy (+0.8% absolute) on Kinetics-400 compared to baseline training. Code is available online.

In this paper, we bring together two divergent strands of research: photometric face capture and statistical 3D face appearance modelling. We propose a novel lightstage capture and processing pipeline for acquiring ear-to-ear, truly intrinsic diffuse and specular albedo maps that fully factor out the effects of illumination, camera and geometry. Using this pipeline, we capture a dataset of 50 scans and combine them with the only existing publicly available albedo dataset (3DRFE) of 23 scans. This allows us to build the first morphable face albedo model. We believe this is the first statistical analysis of the variability of facial specular albedo maps. This model can be used as a plug in replacement for the texture model of the Basel Face Model and we make our new albedo model publicly available. We ensure careful spectral calibration such that our model is built in a linear sRGB space, suitable for inverse rendering of images taken by typical cameras. We demonstrate our model in a state of the art analysis-by-synthesis 3DMM fitting pipeline, are the first to integrate specular map estimation and outperform the Basel Face Model in albedo reconstruction.

Under the conventional diffuse shading model with unknown directional lighting, the set of quadratic surface shapes that are consistent with the spatial derivatives of intensity at a single image point is a two-dimensional algebraic variety embedded in the five-dimensional space of quadratic shapes. We describe the geometry of this variety, and we introduce a concise feedforward model that computes an explicit, differentiable approximation of the variety from the intensity and its derivatives at any single image point. The result is a parallelizable processor that operates at each image point and produces a lighting-invariant descriptor of the continuous set of compatible surface shapes at the point. We describe two applications of this processor: two-shot uncalibrated photometric stereo and quadratic-surface shape from shading.

Implicit Generative Models (IGMs) such as GANs have emerged as effective data-driven models for generating samples, particularly images. In this paper, we formulate the problem of learning an IGM as minimizing the expected distance between characteristic functions. Specifically, we minimize the distance between characteristic functions of the real and generated data distributions under a suitably-chosen weighting distribution. This distance metric, which we term as the characteristic function distance (CFD), can be (approximately) computed with linear time-complexity in the number of samples, in contrast with the quadratic-time Maximum Mean Discrepancy (MMD). By replacing the discrepancy measure in the critic of a GAN with the CFD, we obtain a model that is simple to implement and stable to train. The proposed metric enjoys desirable theoretical properties including continuity and differentiability with respect to generator parameters, and continuity in the weak topology. We further propose a variation of the CFD in which the weighting distribution parameters are also optimized during training; this obviates the need for manual tuning, and leads to an improvement in test power relative to CFD. We demonstrate experimentally that our proposed method outperforms WGAN and MMD-GAN variants on a variety of unsupervised image generation benchmarks.

his paper contributes a novel realtime multi-person motion capture algorithm using multiview video inputs. Due to the heavy occlusions and closely interacting motions in each view, joint optimization on the multiview images and multiple temporal frames is indispensable, which brings up the essential challenge of realtime efficiency. To this end, for the first time, we unify per-view parsing, cross-view matching, and temporal tracking into a single optimization framework, i.e., a 4D association graph that each dimension (image space, viewpoint and time) can be treated equally and simultaneously. To solve the 4D association graph efficiently, we further contribute the idea of 4D limb bundle parsing based on heuristic searching, followed with limb bundle assembling by proposing a bundle Kruskal's algorithm. Our method enables a realtime motion capture system running at 30fps using 5 cameras on a 5-person scene. Benefiting from the unified parsing, matching and tracking constraints, our method is robust to noisy detection due to severe occlusions and close interacting motions, and achieves high-quality online pose reconstruction quality. The proposed method outperforms state-of-the-art methods quantitatively without using high-level appearance information.

Current supervised methods for facial landmark detection require a large amount of training data and may suffer from overfitting to specific datasets due to the massive number of parameters. We introduce a semi-supervised method in which the crucial idea is to first generate implicit face knowledge from the large amounts of unlabeled images of faces available today. In a first, completely unsupervised stage, we train an adversarial autoencoder to reconstruct faces via a low-dimensional face embedding. In a second, supervised stage, we interleave the decoder with transfer layers to retask the generation of color images to the prediction of landmark heatmaps. Our framework (3FabRec) achieves state-of-the-art performance on several common benchmarks and, most importantly, is able to maintain impressive accuracy on extremely small training sets down to as few as 10 images. As the interleaved layers only add a low amount of parameters to the decoder, inference runs at several hundred FPS on a GPU.

Prevalence of voxel-based 3D single-stage detectors contrast with underexplored point-based methods. In this paper, we present a lightweight point-based 3D single stage object detector 3DSSD to achieve decent balance of accuracy and efficiency. In this paradigm, all upsampling layers and the refinement stage, which are indispensable in all existing point-based methods, are abandoned. We instead propose a fusion sampling strategy in downsampling process to make detection on less representative points feasible. A delicate box prediction network, including a candidate generation layer and an anchor-free regression head with a 3D center-ness assignment strategy, is developed to meet the demand of high accuracy and speed. Our 3DSSD paradigm is an elegant single-stage anchor-free one. We evaluate it on widely used KITTI dataset and more challenging nuScenes dataset. Our method outperforms all state-of-the-art voxel-based single-stage methods by a large margin, and even yields comparable performance with two-stage point-based methods, with amazing inference speed of 25+ FPS, 2x faster than former state-of-the-art point-based methods.

Although cameras are ubiquitous, robotic platforms typically rely on active sensors like LiDAR for direct 3D perception. In this work, we propose a novel self-supervised monocular depth estimation method combining geometry with a new deep network, PackNet, learned only from unlabeled monocular videos. Our architecture leverages novel symmetrical packing and unpacking blocks to jointly learn to compress and decompress detail-preserving representations using 3D convolutions. Although self-supervised, our method outperforms other self, semi, and fully supervised methods on the KITTI benchmark. The 3D inductive bias in PackNet enables it to scale with input resolution and number of parameters without overfitting, generalizing better on out-of-domain data such as the NuScenes dataset. Furthermore, it does not require large-scale supervised pretraining on ImageNet and can run in real-time. Finally, we release DDAD (Dense Depth for Automated Driving), a new urban driving dataset with more challenging and accurate depth evaluation, thanks to longer-range and denser ground-truth depth generated from high-density LiDARs mounted on a fleet of self-driving cars operating world-wide.