CVPR 2021 Papers

We introduce KeypointDeformer, a novel unsupervised method for shape control through automatically discovered 3D keypoints. We cast this as the problem of aligning a source 3D object to a target 3D object from the same object category. Our method analyzes the difference between the shapes of the two objects by comparing their latent representations. This latent representation is in the form of 3D keypoints that are learned in an unsupervised way. The difference between the 3D keypoints of the source and the target objects then informs the shape deformation algorithm that deforms the source object into the target object. The whole model is learned end-to-end and simultaneously discovers 3D keypoints while learning to use them for deforming object shapes. Our approach produces intuitive and semantically consistent control of shape deformations. Moreover, our discovered 3D keypoints are consistent across object category instances despite large shape variations. As our method is unsupervised, it can be readily deployed to new object categories without requiring annotations for 3D keypoints and deformations. Project page: http://tomasjakab.github.io/KeypointDeformer

Many recent 6D pose estimation methods exploited object 3D models to generate synthetic images for training because labels come for free. However, due to the domain shift of data distributions between real images and synthetic images, the network trained only on synthetic images fails to capture robust features in real images for 6D pose estimation. We propose to solve this problem by making the network insensitive to different domains, rather than taking the more difficult route of forcing synthetic images to be similar to real images. Inspired by domain adaption methods, a Domain Adaptive Keypoints Detection Network (DAKDN) including a domain adaption layer is used to minimize the discrepancy of deep features between synthetic and real images. A unique challenge here is the lack of ground truth labels (i.e., keypoints) for real images. Fortunately, the geometry relations between keypoints are invariant under real/synthetic domains. Hence, we propose to use the domain-invariant geometry structure among keypoints as a "bridge" constraint to optimize DAKDN for 6D pose estimation across domains. Specifically, DAKDN employs a Graph Convolutional Network (GCN) block to learn the geometry structure from synthetic images and uses the GCN to guide the training for real images. The 6D poses of objects are calculated using Perspective-n-Point (PnP) algorithm based on the predicted keypoints. Experiments show that our method outperforms state-of-the-art approaches without manual poses labels and competes with approaches using manual poses labels.

Data augmentation (DA) is an essential technique for training state-of-the-art deep learning systems. In this paper, we empirically show data augmentation might introduce noisy augmented examples and consequently hurt the performance on unaugmented data during inference. To alleviate this issue, we propose a simple yet highly effective approach, dubbed KeepAugment, to increase augmented images fidelity. The idea is first to use the saliency map to detect important regions on the original images and then preserve these informative regions during augmentation. This information-preserving strategy allows us to generate more faithful training examples. Empirically, we demonstrate our method significantly improves on a number of prior art data augmentation schemes, e.g. AutoAugment, Cutout, random erasing, achieving promising results on image classification, semi-supervised image classification, multi-view multi-camera tracking and object detection.

Modern lane detection methods have achieved remarkable performances in complex real-world scenarios, but many have issues maintaining real-time efficiency, which is important for autonomous vehicles. In this work, we propose LaneATT: an anchor-based deep lane detection model, which, akin to other generic deep object detectors, uses the anchors for the feature pooling step. Since lanes follow a regular pattern and are highly correlated, we hypothesize that in some cases global information may be crucial to infer their positions, especially in conditions such as occlusion, missing lane markers, and others. Thus, this work proposes a novel anchor-based attention mechanism that aggregates global information. The model was evaluated extensively on three of the most widely used datasets in the literature. The results show that our method outperforms the current state-of-the-art methods showing both higher efficacy and efficiency. Moreover, an ablation study is performed along with a discussion on efficiency trade-off options that are useful in practice. Code and models are available at https://github.com/lucastabelini/LaneATT.

We present a new vision-language (VL) pre-training model dubbed Kaleido-BERT, which introduces a novel kaleido strategy for fashion cross-modality representations from transformers. In contrast to random masking strategy of recent VL models, we design alignment guided masking to jointly focus more on image-text semantic relations. To this end, we carry out five novel tasks, i.e., rotation, jigsaw, camouflage, grey-to-color, and blank-to-color for self-supervised VL pre-training at patches of different scale. Kaleido-BERT is conceptually simple and easy to extend to the existing BERT framework, it attains new state-of-the-art results by large margins on four downstream tasks, including text retrieval (R@1: 4.03% absolute improvement), image retrieval (R@1: 7.13% abs imv.), category recognition (ACC: 3.28% abs imv.), and fashion captioning (Bleu4: 1.2 abs imv.). We validate the efficiency of Kaleido-BERT on a wide range of e-commerical websites, demonstrating its broader potential in real-world applications.

Deep Neural Networks (DNN) could forget the knowledge about earlier tasks when learning new tasks, and this is known as catastrophic forgetting. To learn new task without forgetting, recently, the mask-based learning method (e.g. piggyback ) is proposed to address these issues by learning only a binary element-wise mask, while keeping the backbone model fixed. However, the binary mask has limited modeling capacity for new tasks. A more recent work proposes a compress-grow-based method (CPG) to achieve better accuracy for new tasks by partially training backbone model, but with order-higher training cost, which makes it infeasible to be deployed into popular state-of-the-art edge-/mobile-learning. The primary goal of this work is to simultaneously achieve fast and high-accuracy multi-task adaption in a continual learning setting. Thus motivated, we propose a new training method called Kernel-wise Soft Mask (KSM), which learns a kernel-wise hybrid binary and real-value soft mask for each task. Such a soft mask can be viewed as a superposition of a binary mask and a properly scaled real-value tensor, which offers a richer representation capability without low-level kernel support to meet the objective of low hardware overhead. We validate KSM on multiple benchmark datasets against recent state-of-the-art methods (e.g. Piggyback, Packnet, CPG, etc.), which shows good improvement in both accuracy and training cost.

One of the most challenging question types in VQA is when answering the question requires outside knowledge not present in the image. In this work we study open-domain knowledge, the setting when the knowledge required to answer a question is not given/annotated, neither at training nor test time. We tap into two types of knowledge representations and reasoning. First, implicit knowledge which can be learned effectively from unsupervised language pretraining and supervised training data with transformer-based models. Second, explicit, symbolic knowledge encoded in knowledge bases. Our approach combines both---exploiting the powerful implicit reasoning of transformer models for answer prediction, and integrating symbolic representations from a knowledge graph, while never losing their explicit semantics to an implicit embedding. We combine diverse sources of knowledge to cover the wide variety of knowledge needed to solve knowledge-based questions. We show our approach, KRISP, significantly outperforms state-of-the-art on OK-VQA, the largest available dataset for open-domain knowledge-based VQA. We show with extensive ablations that while our model successfully exploits implicit knowledge reasoning, the symbolic answer module which explicitly connects the knowledge graph to the answer vocabulary is critical to the performance of our method and generalizes to rare answers.

Blind super-resolution (SR) methods aim to generate a high quality high resolution image from a low resolution image containing unknown degradations. However, natural images contain various types and amounts of blur: some may be due to the inherent degradation characteristics of the camera, but some may even be intentional, for aesthetic purposes (e.g. Bokeh effect). In the case of the latter, it becomes highly difficult for SR methods to disentangle the blur to remove, and that to leave as is. In this paper, we propose a novel blind SR framework based on kernel-oriented adaptive local adjustment (KOALA) of SR features, called KOALAnet, which jointly learns spatially-variant degradation and restoration kernels in order to adapt to the spatially-variant blur characteristics in real images. Our KOALAnet outperforms recent blind SR methods for synthesized LR images obtained with randomized degradations, and we further show that the proposed KOALAnet produces the most natural results for artistic photographs with intentional blur, which are not over-sharpened, by effectively handling images mixed with in-focus and out-of-focus areas.

We propose Joint-DetNAS, a unified NAS framework for object detection, which integrates 3 key components: Neural Architecture Search, pruning, and Knowledge Distillation. Instead of naively pipelining these techniques, our Joint-DetNAS optimizes them jointly. The algorithm consists of two core processes: student morphism optimizes the student's architecture and removes the redundant parameters, while dynamic distillation aims to find the optimal matching teacher. For student morphism, weight inheritance strategy is adopted, allowing the student to flexibly update its architecture while fully utilize the predecessor's weights, which considerably accelerates the search; To facilitate dynamic distillation, an elastic teacher pool is trained via integrated progressive shrinking strategy, from which teacher detectors can be sampled without additional cost in subsequent searches. Given a base detector as the input, our algorithm directly outputs the derived student detector with high performance without additional training. Experiments demonstrate that our Joint-DetNAS outperforms the naive pipelining approach by a great margin. Given a classic R101-FPN as the base detector, Joint-DetNAS is able to boost its mAP from 41.4 to 43.9 on MS COCO and reduce the latency by 47%, which is on par with the SOTA EfficientDet while requiring less search cost. We hope our proposed method can provide the community with a new way of jointly optimizing NAS, KD and pruning.

This paper considers the problem of unsupervised person re-identification (re-ID), which aims to learn discriminative models with unlabeled data. One popular method is to obtain pseudo-label by clustering and use them to optimize the model. Although this kind of approach has shown promising accuracy, it is hampered by 1) noisy labels produced by clustering and 2) feature variations caused by camera shift. The former will lead to incorrect optimization and thus hinders the model accuracy. The latter will result in assigning the intra-class samples of different cameras to different pseudo-label, making the model sensitive to camera variations. In this paper, we propose a unified framework to solve both problems. Concretely, we propose a Dynamic and Symmetric Cross-Entropy loss (DSCE) to deal with noisy samples and a camera-aware meta-learning algorithm (MetaCam) to adapt camera shift. DSCE can alleviate the negative effects of noisy samples and accommodate the change of clusters after each clustering step. MetaCam simulates cross-camera constraint by splitting the training data into meta-train and meta-test based on camera IDs. With the interacted gradient from meta-train and meta-test, the model is enforced to learn camera-invariant features. Extensive experiments on three re-ID benchmarks show the effectiveness and the complementary of the proposed DSCE and MetaCam. Our method outperforms the state-of-the-art methods on both fully unsupervised re-ID and unsupervised domain adaptive re-ID.

Training of Convolutional Neural Networks (CNNs) with data with noisy labels is known to be a challenge. Based on the fact that directly providing the label to the data (Positive Learning; PL) has a risk of allowing CNNs to memorize the contaminated labels for the case of noisy data, the indirect learning approach that uses complementary labels (Negative Learning for Noisy Labels; NLNL) has proven to be highly effective in preventing overfitting to noisy data as it reduces the risk of providing faulty target. NLNL further employs a three-stage pipeline to improve convergence. As a result, filtering noisy data through the NLNL pipeline is cumbersome, increasing the training cost. In this study, we propose a novel improvement of NLNL, named Joint Negative and Positive Learning (JNPL), that unifies the filtering pipeline into a single stage. JNPL trains CNN via two losses, NL+ and PL+, which are improved upon NL and PL loss functions, respectively. We analyze the fundamental issue of NL loss function and develop new NL+ loss function producing gradient that enhances the convergence of noisy data. Furthermore, PL+ loss function is designed to enable faster convergence to expected-to-be-clean data. We show that the NL+ and PL+ train CNN simultaneously, significantly simplifying the pipeline, allowing greater ease of practical use compared to NLNL. With a simple semi-supervised training technique, our method achieves state-of-the-art accuracy for noisy data classification based on the superior filtering ability.

We propose a novel technique for producing high-quality 3D models that match a given target object image or scan. Our method is based on retrieving an existing shape from a database of 3D models and then deforming its parts to match the target shape. Unlike previous approaches that independently focus on either shape retrieval or deformation, we propose a joint learning procedure that simultaneously trains the neural deformation module along with the embedding space used by the retrieval module. This enables our network to learn a deformation-aware embedding space, so that retrieved models are more amenable to match the target after an appropriate deformation. In fact, we use the embedding space to guide the shape pairs used to train the deformation module, so that it invests its capacity in learning deformations between meaningful shape pairs. Furthermore, our novel part-aware deformation module can work with inconsistent and diverse part-structures on the source shapes. We demonstrate the benefits of our joint training not only on our novel framework, but also on other state-of-the-art neural deformation modules proposed in recent years. Lastly, we also show that our jointly-trained method outperforms various non-joint baselines.

Recent self-supervised contrastive learning provides an effective approach for unsupervised person re-identification (ReID) by learning invariance from different views (transformed versions) of an input. In this paper, we incorporate a Generative Adversarial Network (GAN) and a contrastive learning module into one joint training framework. While the GAN provides online data augmentation for contrastive learning, the contrastive module learns view-invariant features for generation. In this context, we propose a mesh-based view generator. Specifically, mesh projections serve as references towards generating novel views of a person. In addition, we propose a view-invariant loss to facilitate contrastive learning between original and generated views. Deviating from previous GAN-based unsupervised ReID methods involving domain adaptation, we do not rely on a labeled source dataset, which makes our method more flexible. Extensive experimental results show that our method significantly outperforms state-of-the-art methods under both, fully unsupervised and unsupervised domain adaptive settings on several large scale ReID datsets.

Magnetic resonance imaging (MRI) can provide diagnostic information with high-resolution and high-contrast images. However, MRI requires a relatively long scan time compared to other medical imaging techniques, where long scan time might occur patient's discomfort and limit the increase in resolution of magnetic resonance (MR) image. In this study, we propose a Joint Deep Model-based MR Image and Coil Sensitivity Reconstruction Network, called Joint-ICNet, which jointly reconstructs an MR image and coil sensitivity maps from undersampled multi-coil k-space data using deep learning networks combined with MR physical models. Joint-ICNet has two main blocks, where one is an MR image reconstruction block that reconstructs an MR image from undersampled multi-coil k-space data and the other is a coil sensitivity maps reconstruction block that estimates coil sensitivity maps from undersampled multi-coil k-space data. The desired MR image and coil sensitivity maps can be obtained by sequentially estimating them with two blocks based on the unrolled network architecture. To demonstrate the performance of Joint-ICNet, we performed experiments with a fastMRI brain dataset for two reduction factors (R = 4 and 8). With qualitative and quantitative results, we demonstrate that our proposed Joint-ICNet outperforms conventional parallel imaging and deep-learning-based methods in reconstructing MR images from undersampled multi-coil k-space data.

Due to the memorization effect in Deep Neural Networks (DNNs), training with noisy labels usually results in inferior model performance. Existing state-of-the-art methods primarily adopt a sample selection strategy, which selects small-loss samples for subsequent training. However, prior literature tends to perform sample selection within each mini-batch, neglecting the imbalance of noise ratios in different mini-batches. Moreover, valuable knowledge within high-loss samples is wasted. To this end, we propose a noise-robust approach named Jo-SRC (Joint Sample Selection and Model Regularization based on Consistency). Specifically, we train the network in a contrastive learning manner. Predictions from two different views of each sample are used to estimate its "likelihood" of being clean or out-of-distribution. Furthermore, we propose a joint loss to advance the model generalization performance by introducing consistency regularization. Extensive experiments and ablation studies have validated the superiority of our approach over existing state-of-the-art methods.

Unsupervised representation learning with contrastive learning achieves great success recently. However, these methods have to duplicate each training batch to construct contrastive pairs, ie, each training batch and its augmented version should be forwarded simultaneously, leading to nearly double computation resource demand. We propose a novel Jigsaw Clustering pretext task in this paper, which only needs to forward each training batch itself, nearly reducing the training cost by a half. Our method makes use of information from both intra-image and inter-images, and outperforms previous single-batch based methods by a large margin, even comparable to the costly contrastive learning methods with only half the number of training batches. Our method shows that multiple batches during training are not necessary, and opens a new door for future research of single-batch based unsupervised methods. Our models trained on ImageNet datasets achieve state-of-the-art results with linear classification, outperform previous single-batch methods by 2.6%. Models transfer to COCO datasets outperforms MoCo v2 by 0.4% with only half the number of training samples. Our pretrained models outperform supervised ImageNet pretrained models on CIFAR-10 and CIFAR-100 datasets by 0.9% and 4.1% respectively.

In this paper, we are tackling the proposal-free referring expression grounding task, aiming at localizing the target object according to a query sentence, without relying on off-the-shelf object proposals. Existing proposal-free methods employ a query-image matching branch to select the highest-score point in the image feature map as the target box center, with its width and height predicted by another branch. Such methods, however, fail to utilize the contextual relation between the target and reference objects, and lack interpretability on its reasoning procedure. To solve these problems, we propose an iterative shrinking mechanism to localize the target, where the shrinking direction is decided by a reinforcement learning agent, with all contents within the current image patch comprehensively considered. Beside, the sequential shrinking process enables to demonstrate the reasoning about how to iteratively find the target. Experiments show that the proposed method boosts the accuracy by 4.32% against the previous state-of-the-art (SOTA) method on the RefCOCOg dataset, where query sentences are long and complex, with many targets referred by other reference objects.

We propose a novel end-to-end learning-based approach for single image defocus deblurring. The proposed approach is equipped with a novel Iterative Filter Adaptive Network (IFAN) that is specifically designed to handle spatially-varying and large defocus blur. For adaptively handling spatially-varying blur, IFAN predicts pixel-wise deblurring filters, which are applied to defocused features of an input image to generate deblurred features. For effectively managing large blur, IFAN models deblurring filters as stacks of small-sized separable filters. Predicted separable deblurring filters are applied to defocused features using a novel Iterative Adaptive Convolution (IAC) layer. We also propose a training scheme based on defocus disparity estimation and reblurring, which significantly boosts the deblurring quality. We demonstrate that our method achieves state-of-the-art performance both quantitatively and qualitatively on real-world images.

Finding correspondences between shapes is a fundamental problem in computer vision and graphics, which is relevant for many applications, including 3D reconstruction, object tracking, and style transfer. The vast majority of correspondence methods aim to find a solution between pairs of shapes, even if multiple instances of the same class are available. While isometries are often studied in shape correspondence problems, they have not been considered explicitly in the multi-matching setting. This paper closes this gap by proposing a novel optimisation formulation for isometric multi-shape matching. We present a suitable optimisation algorithm for solving our formulation and provide a convergence and complexity analysis. Moreover, our algorithm obtains multi-matchings that are cycle-consistent without having to explicitly enforce cycle-consistency constraints. We demonstrate the superior performance of our method on various datasets and set the new state-of-the-art in isometric multi-shape matching.