ICCV 2021 Papers

Neural Radiance Fields (NeRF) have recently gained a surge of interest within the computer vision community for its power to synthesize photorealistic novel views of real-world scenes. One limitation of NeRF, however, is its requirement of known camera poses to learn the scene representations. In this paper, we propose Bundle-Adjusting Neural Radiance Fields (BARF) for training NeRF from imperfect camera poses -- the joint problem of learning neural 3D representations and registering camera frames. We establish a theoretical connection to classical planar image registration and show that coarse-to-fine registration is also applicable to NeRF. Furthermore, we demonstrate mathematically that positional encoding has a direct impact on the basin of attraction for registration with a synthesis-based objective. Experiments on synthetic and real-world data show that BARF can effectively optimize the neural scene representations and resolve large camera pose misalignment at the same time. This enables applications of view synthesis and localization of video sequences from unknown camera poses, opening up new avenues for visual localization systems (e.g. SLAM) towards sequential registration with NeRF.

Existing cross-domain semantic segmentation methods usually focus on the overall segmentation results of whole objects but neglect the importance of object boundaries. In this work, we find that the segmentation performance can be considerably boosted if we treat object boundaries properly. For that, we propose a novel method called BAPA-Net, which is based on a convolutional neural network via Boundary Adaptation and Prototype Alignment, under the unsupervised domain adaptation setting. Specifically, we first construct additional images by pasting objects from source images to target images, and we develop a so-called boundary adaptation module to weigh each pixel based on its distance to the nearest boundary pixel of those pasted source objects. Moreover, we pro- pose another prototype alignment module to reduce the domain mismatch by minimizing distances between the class prototypes of the source and target domains, where boundaries are removed to avoid domain confusion during prototype calculation. By integrating the boundary adaptation and prototype alignment, we are able to train a discriminative and domain-invariant model for cross-domain semantic segmentation. We conduct extensive experiments on the benchmark datasets of urban scenes (i.e., GTA5->Cityscapes and SYNTHIA->Cityscapes). And the promising results clearly show the effectiveness of our BAPA-Net method over existing state-of-the-art for cross-domain semantic segmentation. Our implementation is available at https://github.com/manmanjun/BAPA-Net.

ObjectGoal Navigation (ObjectNav) is an embodied task wherein agents are to navigate to an object instance in an unseen environment. Prior works have shown that end-to-end ObjectNav agents that use vanilla visual and recurrent modules, e.g. a CNN+RNN, perform poorly due to overfitting and sample inefficiency. This has motivated current state-of-the-art methods to mix analytic and learned components and operate on explicit spatial maps of the environment. We instead re-enable a generic learned agent by adding auxiliary learning tasks and an exploration reward. Our agents achieve 24.5% success and 8.1% SPL, a 37% and 8% relative improvement over prior state-of-the-art, respectively, on the Habitat ObjectNav Challenge. From our analysis, we propose that agents will act to simplify their visual inputs so as to smooth their RNN dynamics, and that auxiliary tasks reduce overfitting by minimizing effective RNN dimensionality; i.e. a performant ObjectNav agent that must maintain coherent plans over long horizons does so by learning smooth, low-dimensional recurrent dynamics.

Current neural architecture search (NAS) algorithms still require expert knowledge and effort to design a search space for network construction. In this paper, we consider automating the search space design to minimize human interference, which however faces two challenges: the explosive complexity of the exploration space and the expensive computation cost to evaluate the quality of different search spaces. To solve them, we propose a novel differentiable evolutionary framework named AutoSpace, which evolves the search space to an optimal one with following novel techniques: a differentiable fitness scoring function to efficiently evaluate the performance of cells and a reference architecture to speedup the evolution procedure and avoid falling into sub-optimal solutions. The framework is generic and compatible with additional computational constraints, making it feasible to learn specialized search spaces that fit different computational budgets. With the learned search space, the performance of recent NAS algorithms can be improved significantly compared with using manually de-signed spaces. Remarkably, the models generated from the new search space achieve 77.8% top-1 accuracy on ImageNet under the mobile setting (MAdds<=500M), outperforming previous SOTA EfficientNet-B0 by 0.7%. https://github.com/zhoudaquan/AutoSpace.git

Existing deep learning-based approaches for monocular 3D object detection in autonomous driving often model the object as a rotated 3D cuboid while the object's geometric shape has been ignored. In this work, we propose an approach for incorporating the shape-aware 2D/3D constraints into the 3D detection framework. Specifically, we employ the deep neural network to learn distinguished 2D keypoints in the 2D image domain and regress their corresponding 3D coordinates in the local 3D object coordinate first. Then the 2D/3D geometric constraints are built by these correspondences for each object to boost the detection performance. For generating the ground truth of 2D/3D keypoints, an automatic model-fitting approach has been proposed by fitting the deformed 3D object model and the object mask in the 2D image. The proposed framework has been verified on the public KITTI dataset and the experimental results demonstrate that by using additional geometrical constraints the detection performance has been significantly improved as compared to the baseline method. More importantly, the proposed framework achieves state-of-the-art performance with real time. Data and code will be available at https://github.com/zongdai/AutoShape

Recently, pure transformer-based models have shown great potentials for vision tasks such as image classification and detection. However, the design of transformer networks is challenging. It has been observed that the depth, embedding dimension, and number of heads can largely affect the performance of vision transformers. Previous models configure these dimensions based upon manual crafting. In this work, we propose a new one-shot architecture search framework, namely AutoFormer, dedicated to vision transformer search. AutoFormer entangles the weights of different blocks in the same layers during supernet training. Benefiting from the strategy, the trained supernet allows thousands of subnets to be very well-trained. Specifically, the performance of these subnets with weights inherited from the supernet is comparable to those retrained from scratch. Besides, the searched models, which we refer to AutoFormers, surpass the recent state-of-the-arts such as ViT and DeiT. In particular, AutoFormer-tiny/small/base achieve 74.7%/81.7%/82.4% top-1 accuracy on ImageNet with 5.7M/22.9M/53.7M parameters, respectively. Lastly, we verify the transferability of AutoFormer by providing the performance on downstream benchmarks and distillation experiments. Code and models are available at https://github.com/microsoft/Cream.

We propose an Auto-Parsing Network (APN) to discover and exploit the input data's hidden tree structures for improving the effectiveness of the Transformer-based vision-language systems. Specifically, we impose a Probabilistic Graphical Model (PGM) parameterized by the attention operations on each self-attention layer to incorporate sparse assumption. We use this PGM to softly segment an input sequence into a few clusters where each cluster can be treated as the parent of the inside entities. By stacking these PGM constrained self-attention layers, the clusters in a lower layer compose into a new sequence, and the PGM in a higher layer will further segment this sequence. Iteratively, a sparse tree can be implicitly parsed, and this tree's hierarchical knowledge is incorporated into the transformed embeddings, which can be used for solving the target vision-language tasks. Specifically, we showcase that our APN can strengthen Transformer based networks in two major vision-language tasks: Captioning and Visual Question Answering. Also, a PGM probability-based parsing algorithm is developed by which we can discover what the hidden structure of input is during the inference.

Model compression aims to deploy deep neural networks (DNN) on mobile devices with limited computing and storage resources. However, most of the existing model compression methods rely on manually defined rules, which require domain expertise. DNNs are essentially computational graphs, which contain rich structural information. In this paper, we aim to find a suitable compression policy from DNNs' structural information. We propose an automatic graph encoder-decoder model compression (AGMC) method combined with graph neural networks (GNN) and reinforcement learning (RL). We model the target DNN as a graph and use GNN to learn the DNN's embeddings automatically. We compared our method with rule-based DNN embedding model compression methods to show the effectiveness of our method. Results show that our learning-based DNN embedding achieves better performance and a higher compression ratio with fewer search steps. We evaluated our method on over-parameterized and mobile-friendly DNNs and compared our method with handcrafted and learning-based model compression approaches. On over parameterized DNNs, such as ResNet-56, our method outperformed handcrafted and learning-based methods with 4.36% and 2.56% higher accuracy, respectively. Furthermore, on MobileNet-v2, we achieved a higher compression ratio than state-of-the-art methods with just 0.93% accuracy loss.

Modern cameras are equipped with a wide array of sensors that enable recording the geospatial context of an image. Taking advantage of this, we explore depth estimation under the assumption that the camera is geocalibrated, a problem we refer to as geo-enabled depth estimation. Our key insight is that if capture location is known, the corresponding overhead viewpoint offers a valuable resource for understanding the scale of the scene. We propose an end-to-end architecture for depth estimation that uses geospatial context to infer a synthetic ground-level depth map from a co-located overhead image, then fuses it inside of an encoder/decoder style segmentation network. To support evaluation of our methods, we extend a recently released dataset with overhead imagery and corresponding height maps. Results demonstrate that integrating geospatial context significantly reduces error compared to baselines, both at close ranges and when evaluating at much larger distances than existing benchmarks consider.

Adversarial attack algorithms are dominated by penalty methods, which are slow in practice, or more efficient distance-customized methods, which are heavily tailored to the properties of the considered distance. We propose a white-box attack algorithm to generate minimally perturbed adversarial examples based on Augmented Lagrangian principles. We bring several algorithmic modifications, which have a crucial effect on performance. Our attack enjoys the generality of penalty methods and the computational efficiency of distance-customized algorithms, and can be readily used for a wide set of distances. We compare our attack to state-of-the-art methods on three datasets and several models, and consistently obtain competitive performances with similar or lower computational complexity.

Generating conversational gestures from speech audio is challenging due to the inherent one-to-many mapping between audio and body motions. Conventional CNNs/RNNs assume one-to-one mapping, and thus tend to predict the average of all possible target motions, resulting in plain/boring motions during inference. In order to overcome this problem, we propose a novel conditional variational autoencoder (VAE) that explicitly models one-to-many audio-to-motion mapping by splitting the cross-modal latent code into shared code and motion-specific code. The shared code mainly models the strong correlation between audio and motion (such as the synchronized audio and motion beats), while the motion-specific code captures diverse motion information independent of the audio. However, splitting the latent code into two parts poses training difficulties for the VAE model. A mapping network facilitating random sampling along with other techniques including relaxed motion loss, bicycle constraint, and diversity loss are designed to better train the VAE. Experiments on both 3D and 2D motion datasets verify that our method generates more realistic and diverse motions than state-of-the-art methods, quantitatively and qualitatively. Finally, we demonstrate that our method can be readily used to generate motion sequences with user-specified motion clips on the timeline. Code and more results are at https://jingli513.github.io/audio2gestures.

Given only a few glimpses of an environment, how much can we infer about its entire floorplan? Existing methods can map only what is visible or immediately apparent from context, and thus require substantial movements through a space to fully map it. We explore how both audio and visual sensing together can provide rapid floorplan reconstruction from limited viewpoints. Audio not only helps sense geometry outside the camera's field of view, but it also reveals the existence of distant freespace (e.g., a dog barking in another room) and suggests the presence of rooms not visible to the camera (e.g., a dishwasher humming in what must be the kitchen to the left). We introduce AV-Map, a novel multi-modal encoder-decoder framework that reasons jointly about audio and vision to reconstruct a floorplan from a short input video sequence. We train our model to predict both the interior structure of the environment and the associated rooms' semantic labels. Our results on 85 large real-world environments show the impact: with just a few glimpses spanning 26% of an area, we can estimate the whole area with 66% accuracy---substantially better than the state of the art approach for extrapolating visual maps.

Estimating the motion of the camera together with the 3D structure of the scene from a monocular vision system is a complex task that often relies on the so-called scene rigidity assumption. When observing a dynamic environment, this assumption is violated which leads to an ambiguity between the ego-motion of the camera and the motion of the objects. To solve this problem, we present a self-supervised learning framework for 3D object motion field estimation from monocular videos. Our contributions are two-fold. First, we propose a two-stage projection pipeline to explicitly disentangle the camera ego-motion and the object motions with dynamics attention module, called DAM. Specifically, we design an integrated motion model that estimates the motion of the camera and object in the first and second warping stages, respectively, controlled by the attention module through a shared motion encoder. Second, we propose an object motion field estimation through contrastive sample consensus, called CSAC, taking advantage of weak semantic prior (bounding box from an object detector) and geometric constraints (each object respects the rigid body motion model). Experiments on KITTI, Cityscapes, and Waymo Open Dataset demonstrate the relevance of our approach and show that our method outperforms state-of-the-art algorithms for the tasks of self-supervised monocular depth estimation, object motion segmentation, monocular scene flow estimation, and visual odometry.

The core of visual place recognition (VPR) lies in how to identify task-relevant visual cues and embed them into discriminative representations. Focusing on these two points, we propose a novel encoding strategy named Attentional Pyramid Pooling of Salient Visual Residuals (APPSVR). It incorporates three types of attention modules to model the saliency of local features in individual, spatial and cluster dimensions respectively. (1) To inhibit task-irrelevant local features, a semantic-reinforced local weighting scheme is employed for local feature refinement; (2) To leverage the spatial context, an attentional pyramid structure is constructed to adaptively encode regional features according to their relative spatial saliency; (3) To distinguish the different importance of visual clusters to the task, a parametric normalization is proposed to adjust their contribution to image descriptor generation. Experiments demonstrate APPSVR outperforms the existing techniques and achieves a new state-of-the-art performance on VPR benchmark datasets. The visualization shows the saliency map learned in a weakly supervised manner is largely consistent with human cognition.

This paper proposes a novel Attention-based Multi-Reference Super-resolution network (AMRSR) that, given a low-resolution image, learns to adaptively transfer the most similar texture from multiple reference images to the super-resolution output whilst maintaining spatial coherence. The use of multiple reference images together with attention-based sampling is demonstrated to achieve significantly improved performance over state-of-the-art reference super-resolution approaches on multiple benchmark datasets. Reference super-resolution approaches have recently been proposed to overcome the ill-posed problem of image super-resolution by providing additional information from a high-resolution reference image. Multi-reference super-resolution extends this approach by providing a more diverse pool of image features to overcome the inherent information deficit whilst maintaining memory efficiency. A novel hierarchical attention-based sampling approach is introduced to learn the similarity between low-resolution image features and multiple reference images based on a perceptual loss. Ablation demonstrates the contribution of both multi-reference and hierarchical attention-based sampling to overall performance. Perceptual and quantitative ground-truth evaluation demonstrates significant improvement in performance even when the reference images deviate significantly from the target image. The project website can be found at https://marcopesavento.github.io/AMRSR/

Videos flow as the mixture of language, acoustic, and vision modalities. A thorough video understanding needs to fuse time-series data of different modalities for prediction. Due to the variable receiving frequency for sequences from each modality, there usually exists inherent asynchrony across the collected multimodal streams. Towards an efficient multimodal fusion from asynchronous multimodal streams, we need to model the correlations between elements from different modalities. The recent Multimodal Transformer (MulT) approach extends the self-attention mechanism of the original Transformer network to learn the crossmodal dependencies between elements. However, the direct replication of self-attention will suffer from the distribution mismatch across different modality features. As a result, the learnt crossmodal dependencies can be unreliable. Motivated by this observation, this work proposes the Modality-Invariant Crossmodal Attention (MICA) approach towards learning crossmodal interactions over modality-invariant space in which the distribution mismatch between different modalities is well bridged. To this end, both the marginal distribution and the elements with high-confidence correlations are aligned over the common space of the query and key vectors which are computed from different modalities. Experiments on three standard benchmarks of multimodal video understanding clearly validate the superiority of our approach.

In unconstrained real-world surveillance scenarios, person re-identification (Re-ID) models usually suffer from different low-level perceptual variations, e.g., cross-resolution and insufficient lighting. Due to the limited variation range of training data, existing models are difficult to generalize to scenes with unknown perceptual interference types. To address the above problem, in this paper, we propose two disjoint data-generation ways to complement existing training samples to improve the robustness of Re-ID models. Firstly, considering the sparsity and imbalance of samples in the perceptual space, a dense resampling method from the estimated perceptual distribution is performed. Secondly, to dig more representative generated samples for identity representation learning, we introduce a graph-based white-box attacker to guide the data generation process with intra-batch ranking and discriminate attention. In addition, two synthetic-to-real feature constraints are introduced into the Re-ID training to prevent the generated data from bringing domain bias. Our method is effective, easy-to-implement, and independent of the specific network architecture. Applying our approach to a ResNet-50 baseline can already achieve competitive results, surpassing state-of-the-art methods by +1.2% at Rank-1 on the MLR-CUHK03 dataset.

Due to the great success of image-to-image (Img2Img) translation GANs, many applications with ethics issues arise, e.g., DeepFake and DeepNude, presenting a challenging problem to prevent the misuse of these techniques. In this work, we tackle the problem by a new adversarial attack scheme, namely the Nullifying Attack, which cancels the image translation process and proposes a corresponding framework, the Limit-Aware Self-Guiding Gradient Sliding Attack (LaS-GSA) under a black-box setting. In other words, by processing the image with the proposed LaS-GSA before publishing, any image translation functions can be nullified, which prevents the images from malicious manipulations. First, we introduce the limit-aware RGF and gradient sliding mechanism to estimate the gradient that adheres to the adversarial limit, i.e., the pixel value limitations of the adversarial example. We theoretically prove that our model is able to avoid the error caused by the projection operation in both the direction and the length. Then, an effective self-guiding prior is extracted solely from the threat model and the target image to efficiently leverage the prior information and guide the gradient estimation process. Extensive experiments demonstrate that LaS-GSA requires fewer queries to nullify the image translation process with higher success rates than 4 state-of-the-art methods.

In a typical multi-label setting, a picture contains on average few positive labels, and many negative ones. This positive-negative imbalance dominates the optimization process, and can lead to under-emphasizing gradients from positive labels during training, resulting in poor accuracy. In this paper, we introduce a novel asymmetric loss (""ASL""), which operates differently on positive and negative samples. The loss enables to dynamically down-weights and hard-thresholds easy negative samples, while also discarding possibly mislabeled samples. We demonstrate how ASL can balance the probabilities of different samples, and how this balancing is translated to better mAP scores. With ASL, we reach state-of-the-art results on multiple popular multi-label datasets: MS-COCO, Pascal-VOC, NUS-WIDE and Open Images. We also demonstrate ASL applicability for other tasks, such as single-label classification and object detection. ASL is effective, easy to implement, and does not increase the training time or complexity. Implementation is available at: https://github.com/Alibaba-MIIL/ASL.