CVPR 2023 Papers

Many perception systems in mobile computing, autonomous navigation, and AR/VR face strict compute constraints that are particularly challenging for high-resolution input images. Previous works propose nonuniform downsamplers that "learn to zoom" on salient image regions, reducing compute while retaining task-relevant image information. However, for tasks with spatial labels (such as 2D/3D object detection and semantic segmentation), such distortions may harm performance. In this work (LZU), we "learn to zoom" in on the input image, compute spatial features, and then "unzoom" to revert any deformations. To enable efficient and differentiable unzooming, we approximate the zooming warp with a piecewise bilinear mapping that is invertible. LZU can be applied to any task with 2D spatial input and any model with 2D spatial features, and we demonstrate this versatility by evaluating on a variety of tasks and datasets: object detection on Argoverse-HD, semantic segmentation on Cityscapes, and monocular 3D object detection on nuScenes. Interestingly, we observe boosts in performance even when high-resolution sensor data is unavailable, implying that LZU can be used to "learn to upsample" as well. Code and additional visuals are available at https://tchittesh.github.io/lzu/.

We introduce a method for novel view synthesis given only a single wide-baseline stereo image pair. In this challenging regime, 3D scene points are regularly observed only once, requiring prior-based reconstruction of scene geometry and appearance. We find that existing approaches to novel view synthesis from sparse observations fail due to recovering incorrect 3D geometry and the high cost of differentiable rendering that precludes their scaling to large-scale training. We take a step towards resolving these shortcomings by formulating a multi-view transformer encoder, proposing an efficient, image-space epipolar line sampling scheme to assemble image features for a target ray, and a lightweight cross-attention-based renderer. Our contributions enable training of our method on a large-scale real-world dataset of indoor and outdoor scenes. In several ablation studies, we demonstrate that our contributions enable learning of powerful multi-view geometry priors while reducing both rendering time and memory footprint. We conduct extensive comparisons on held-out test scenes across two real-world datasets, significantly outperforming prior work on novel view synthesis from sparse image observations and achieving multi-view-consistent novel view synthesis.

We introduce a method that can learn to predict scene-level implicit functions for 3D reconstruction from posed RGBD data. At test time, our system maps a previously unseen RGB image to a 3D reconstruction of a scene via implicit functions. While implicit functions for 3D reconstruction have often been tied to meshes, we show that we can train one using only a set of posed RGBD images. This setting may help 3D reconstruction unlock the sea of accelerometer+RGBD data that is coming with new phones. Our system, D2-DRDF, can match and sometimes outperform current methods that use mesh supervision and shows better robustness to sparse data.

Multi-frame depth estimation generally achieves high accuracy relying on the multi-view geometric consistency. When applied in dynamic scenes, e.g., autonomous driving, this consistency is usually violated in the dynamic areas, leading to corrupted estimations. Many multi-frame methods handle dynamic areas by identifying them with explicit masks and compensating the multi-view cues with monocular cues represented as local monocular depth or features. The improvements are limited due to the uncontrolled quality of the masks and the underutilized benefits of the fusion of the two types of cues. In this paper, we propose a novel method to learn to fuse the multi-view and monocular cues encoded as volumes without needing the heuristically crafted masks. As unveiled in our analyses, the multi-view cues capture more accurate geometric information in static areas, and the monocular cues capture more useful contexts in dynamic areas. To let the geometric perception learned from multi-view cues in static areas propagate to the monocular representation in dynamic areas and let monocular cues enhance the representation of multi-view cost volume, we propose a cross-cue fusion (CCF) module, which includes the cross-cue attention (CCA) to encode the spatially non-local relative intra-relations from each source to enhance the representation of the other. Experiments on real-world datasets prove the significant effectiveness and generalization ability of the proposed method.

This paper addresses the challenge of efficiently reconstructing volumetric videos of dynamic humans from sparse multi-view videos. Some recent works represent a dynamic human as a canonical neural radiance field (NeRF) and a motion field, which are learned from input videos through differentiable rendering. But the per-scene optimization generally requires hours. Other generalizable NeRF models leverage learned prior from datasets to reduce the optimization time by only finetuning on new scenes at the cost of visual fidelity. In this paper, we propose a novel method for learning neural volumetric representations of dynamic humans in minutes with competitive visual quality. Specifically, we define a novel part-based voxelized human representation to better distribute the representational power of the network to different human parts. Furthermore, we propose a novel 2D motion parameterization scheme to increase the convergence rate of deformation field learning. Experiments demonstrate that our model can be learned 100 times faster than previous per-scene optimization methods while being competitive in the rendering quality. Training our model on a 512x512 video with 100 frames typically takes about 5 minutes on a single RTX 3090 GPU. The code is available on our project page: https://zju3dv.github.io/instant_nvr

We propose a novel 3D morphable model for complete human heads based on hybrid neural fields. At the core of our model lies a neural parametric representation that disentangles identity and expressions in disjoint latent spaces. To this end, we capture a person's identity in a canonical space as a signed distance field (SDF), and model facial expressions with a neural deformation field. In addition, our representation achieves high-fidelity local detail by introducing an ensemble of local fields centered around facial anchor points. To facilitate generalization, we train our model on a newly-captured dataset of over 3700 head scans from 203 different identities using a custom high-end 3D scanning setup. Our dataset significantly exceeds comparable existing datasets, both with respect to quality and completeness of geometry, averaging around 3.5M mesh faces per scan. Finally, we demonstrate that our approach outperforms state-of-the-art methods in terms of fitting error and reconstruction quality.

In this paper, we propose a novel hybrid representation and end-to-end trainable network architecture to model fully editable and customizable neural avatars. At the core of our work lies a representation that combines the modeling power of neural fields with the ease of use and inherent 3D consistency of skinned meshes. To this end, we construct a trainable feature codebook to store local geometry and texture features on the vertices of a deformable body model, thus exploiting its consistent topology under articulation. This representation is then employed in a generative auto-decoder architecture that admits fitting to unseen scans and sampling of realistic avatars with varied appearances and geometries. Furthermore, our representation allows local editing by swapping local features between 3D assets. To verify our method for avatar creation and editing, we contribute a new high-quality dataset, dubbed CustomHumans, for training and evaluation. Our experiments quantitatively and qualitatively show that our method generates diverse detailed avatars and achieves better model fitting performance compared to state-of-the-art methods. Our code and dataset are available at https://ait.ethz.ch/custom-humans.

Understanding geometric concepts, such as distance and shape, is essential for understanding the real world and also for many vision tasks. To incorporate such information into a visual representation of a scene, we propose learning to represent the scene by sketching, inspired by human behavior. Our method, coined Learning by Sketching (LBS), learns to convert an image into a set of colored strokes that explicitly incorporate the geometric information of the scene in a single inference step without requiring a sketch dataset. A sketch is then generated from the strokes where CLIP-based perceptual loss maintains a semantic similarity between the sketch and the image. We show theoretically that sketching is equivariant with respect to arbitrary affine transformations and thus provably preserves geometric information. Experimental results show that LBS substantially improves the performance of object attribute classification on the unlabeled CLEVR dataset, domain transfer between CLEVR and STL-10 datasets, and for diverse downstream tasks, confirming that LBS provides rich geometric information.

Emotion understanding is an essential but highly challenging component of artificial general intelligence. The absence of extensive annotated datasets has significantly impeded advancements in this field. We present EmotionCLIP, the first pre-training paradigm to extract visual emotion representations from verbal and nonverbal communication using only uncurated data. Compared to numerical labels or descriptions used in previous methods, communication naturally contains emotion information. Furthermore, acquiring emotion representations from communication is more congruent with the human learning process. We guide EmotionCLIP to attend to nonverbal emotion cues through subject-aware context encoding and verbal emotion cues using sentiment-guided contrastive learning. Extensive experiments validate the effectiveness and transferability of EmotionCLIP. Using merely linear-probe evaluation protocol, EmotionCLIP outperforms the state-of-the-art supervised visual emotion recognition methods and rivals many multimodal approaches across various benchmarks. We anticipate that the advent of EmotionCLIP will address the prevailing issue of data scarcity in emotion understanding, thereby fostering progress in related domains. The code and pre-trained models are available at https://github.com/Xeaver/EmotionCLIP.

An image is usually described by more than one attribute like "shape" and "color". When a dataset is biased, i.e., most samples have attributes spuriously correlated with the target label, a Deep Neural Network (DNN) is prone to make predictions by the "unintended" attribute, especially if it is easier to learn. To improve the generalization ability when training on such a biased dataset, we propose a chi^2-model to learn debiased representations. First, we design a chi-shape pattern to match the training dynamics of a DNN and find Intermediate Attribute Samples (IASs) --- samples near the attribute decision boundaries, which indicate how the value of an attribute changes from one extreme to another. Then we rectify the representation with a chi-structured metric learning objective. Conditional interpolation among IASs eliminates the negative effect of peripheral attributes and facilitates retaining the intra-class compactness. Experiments show that chi^2-model learns debiased representation effectively and achieves remarkable improvements on various datasets.

We propose a differentiable nonlinear least squares framework to account for uncertainty in relative pose estimation from feature correspondences. Specifically, we introduce a symmetric version of the probabilistic normal epipolar constraint, and an approach to estimate the covariance of feature positions by differentiating through the camera pose estimation procedure. We evaluate our approach on synthetic, as well as the KITTI and EuRoC real-world datasets. On the synthetic dataset, we confirm that our learned covariances accurately approximate the true noise distribution. In real world experiments, we find that our approach consistently outperforms state-of-the-art non-probabilistic and probabilistic approaches, regardless of the feature extraction algorithm of choice.

Fashion representation learning involves the analysis and understanding of various visual elements at different granularities and the interactions among them. Existing works often learn fine-grained fashion representations at the attribute-level without considering their relationships and inter-dependencies across different classes. In this work, we propose to learn an attribute and class specific fashion representation duet to better model such attribute relationships and inter-dependencies by leveraging prior knowledge about the taxonomy of fashion attributes and classes. Through two sub-networks for the attributes and classes, respectively, our proposed an embedding network progressively learn and refine the visual representation of a fashion image to improve its robustness for fashion retrieval. A multi-granularity loss consisting of attribute-level and class-level losses is proposed to introduce appropriate inductive bias to learn across different granularities of the fashion representations. Experimental results on three benchmark datasets demonstrate the effectiveness of our method, which outperforms the state-of-the-art methods with a large margin.

The goal of this work is to understand the way actions are performed in videos. That is, given a video, we aim to predict an adverb indicating a modification applied to the action (e.g. cut "finely"). We cast this problem as a regression task. We measure textual relationships between verbs and adverbs to generate a regression target representing the action change we aim to learn. We test our approach on a range of datasets and achieve state-of-the-art results on both adverb prediction and antonym classification. Furthermore, we outperform previous work when we lift two commonly assumed conditions: the availability of action labels during testing and the pairing of adverbs as antonyms. Existing datasets for adverb recognition are either noisy, which makes learning difficult, or contain actions whose appearance is not influenced by adverbs, which makes evaluation less reliable. To address this, we collect a new high quality dataset: Adverbs in Recipes (AIR). We focus on instructional recipes videos, curating a set of actions that exhibit meaningful visual changes when performed differently. Videos in AIR are more tightly trimmed and were manually reviewed by multiple annotators to ensure high labelling quality. Results show that models learn better from AIR given its cleaner videos. At the same time, adverb prediction on AIR is challenging, demonstrating that there is considerable room for improvement.

Shape from Polarization (SfP) aims to recover surface normal using the polarization cues of light. The accuracy of existing SfP methods is affected by two main problems. First, the ambiguity of polarization cues partially results in false normal estimation. Second, the widely-used assumption about orthographic projection is too ideal. To solve these problems, we propose the first approach that combines deep learning and stereo polarization information to recover not only normal but also disparity. Specifically, for the ambiguity problem, we design a Shape Consistency-based Mask Prediction (SCMP) module. It exploits the inherent consistency between normal and disparity to identify the areas with false normal estimation. We replace the unreliable features enclosed by these areas with new features extracted by global attention mechanism. As to the orthographic projection problem, we propose a novel Viewing Direction-aided Positional Encoding (VDPE) strategy. This strategy is based on the unique pixel-viewing direction encoding, and thus enables our neural network to handle the non-orthographic projection. In addition, we establish a real-world stereo SfP dataset that contains various object categories and illumination conditions. Experiments showed that compared with existing SfP methods, our approach is more accurate. Moreover, our approach shows higher robustness to light variation.

In this work, we focus on the challenging problem of Label Enhancement (LE), which aims to exactly recover label distributions from logical labels, and present a novel Label Information Bottleneck (LIB) method for LE. For the recovery process of label distributions, the label irrelevant information contained in the dataset may lead to unsatisfactory recovery performance. To address this limitation, we make efforts to excavate the essential label relevant information to improve the recovery performance. Our method formulates the LE problem as the following two joint processes: 1) learning the representation with the essential label relevant information, 2) recovering label distributions based on the learned representation. The label relevant information can be excavated based on the "bottleneck" formed by the learned representation. Significantly, both the label relevant information about the label assignments and the label relevant information about the label gaps can be explored in our method. Evaluation experiments conducted on several benchmark label distribution learning datasets verify the effectiveness and competitiveness of LIB.

Action quality assessment (AQA) has become an emerging topic since it can be extensively applied in numerous scenarios. However, most existing methods and datasets focus on single-person short-sequence scenes, hindering the application of AQA in more complex situations. To address this issue, we construct a new multi-person long-form video dataset for action quality assessment named LOGO. Distinguished in scenario complexity, our dataset contains 200 videos from 26 artistic swimming events with 8 athletes in each sample along with an average duration of 204.2 seconds. As for richness in annotations, LOGO includes formation labels to depict group information of multiple athletes and detailed annotations on action procedures. Furthermore, we propose a simple yet effective method to model relations among athletes and reason about the potential temporal logic in long-form videos. Specifically, we design a group-aware attention module, which can be easily plugged into existing AQA methods, to enrich the clip-wise representations based on contextual group information. To benchmark LOGO, we systematically conduct investigations on the performance of several popular methods in AQA and action segmentation. The results reveal the challenges our dataset brings. Extensive experiments also show that our approach achieves state-of-the-art on the LOGO dataset. The dataset and code will be released at https://github.com/shiyi-zh0408/LOGO.

It is important to quantify the uncertainty of input samples, especially in mission-critical domains such as autonomous driving and healthcare, where failure predictions on out-of-distribution (OOD) data are likely to cause big problems. OOD detection problem fundamentally begins in that the model cannot express what it is not aware of. Post-hoc OOD detection approaches are widely explored because they do not require an additional re-training process which might degrade the model's performance and increase the training cost. In this study, from the perspective of neurons in the deep layer of the model representing high-level features, we introduce a new aspect for analyzing the difference in model outputs between in-distribution data and OOD data. We propose a novel method, Leveraging Important Neurons (LINe), for post-hoc Out of distribution detection. Shapley value-based pruning reduces the effects of noisy outputs by selecting only high-contribution neurons for predicting specific classes of input data and masking the rest. Activation clipping fixes all values above a certain threshold into the same value, allowing LINe to treat all the class-specific features equally and just consider the difference between the number of activated feature differences between in-distribution and OOD data. Comprehensive experiments verify the effectiveness of the proposed method by outperforming state-of-the-art post-hoc OOD detection methods on CIFAR-10, CIFAR-100, and ImageNet datasets.

Humans universally dislike the task of cleaning up a messy room. If machines were to help us with this task, they must understand human criteria for regular arrangements, such as several types of symmetry, co-linearity or co-circularity, spacing uniformity in linear or circular patterns, and further inter-object relationships that relate to style and functionality. Previous approaches for this task relied on human input to explicitly specify goal state, or synthesized scenes from scratch--but such methods do not address the rearrangement of existing messy scenes without providing a goal state. In this paper, we present LEGO-Net, a data-driven transformer-based iterative method for LEarning reGular rearrangement of Objects in messy rooms. LEGO-Net is partly inspired by diffusion models--it starts with an initial messy state and iteratively "de-noises" the position and orientation of objects to a regular state while reducing distance traveled. Given randomly perturbed object positions and orientations in an existing dataset of professionally-arranged scenes, our method is trained to recover a regular re-arrangement. Results demonstrate that our method is able to reliably rearrange room scenes and outperform other methods. We additionally propose a metric for evaluating regularity in room arrangements using number-theoretic machinery.

Language-based colorization produces plausible colors consistent with the language description provided by the user. Recent studies introduce additional annotation to prevent color-object coupling and mismatch issues, but they still have difficulty in distinguishing instances corresponding to the same object words. In this paper, we propose a transformer-based framework to automatically aggregate similar image patches and achieve instance awareness without any additional knowledge. By applying our presented luminance augmentation and counter-color loss to break down the statistical correlation between luminance and color words, our model is driven to synthesize colors with better descriptive consistency. We further collect a dataset to provide distinctive visual characteristics and detailed language descriptions for multiple instances in the same image. Extensive experiments demonstrate our advantages of synthesizing visually pleasing and description-consistent results of instance-aware colorization.