ICCV 2023 论文列表

Supervised learning of image classifiers distills human knowledge into a parametric model through pairs of images and corresponding labels (X,Y). We argue that this simple and widely used representation of human knowledge neglects rich auxiliary information from the annotation procedure, such as the time-series of mouse traces and clicks left after image selection. Our insight is that such annotation byproducts Z provide approximate human attention that weakly guides the model to focus on the foreground cues, reducing spurious correlations and discouraging shortcut learning. To verify this, we create ImageNet-AB and COCO-AB. They are ImageNet and COCO training sets enriched with sample-wise annotation byproducts, collected by replicating the respective original annotation tasks. We refer to the new paradigm of training models with annotation byproducts as learning using annotation byproducts (LUAB). We show that a simple multitask loss for regressing Z together with Y already improves the generalisability and robustness of the learned models. Compared to the original supervised learning, LUAB does not require extra annotation costs. ImageNet-AB and COCO-AB are at https://github.com/naver-ai/NeglectedFreeLunch.

Detecting out-of-distribution (OOD) samples are crucial for machine learning models deployed in open-world environments. Classifier-based scores are a standard approach for OOD detection due to their fine-grained detection capability. However, these scores often suffer from overconfidence issues, misclassifying OOD samples distant from the in-distribution region. To address this challenge, we propose a method called Nearest Neighbor Guidance (NNGuide) that guides the classifier-based score to respect the boundary geometry of the data manifold. NNGuide reduces the overconfidence of OOD samples while preserving the fine-grained capability of the classifier-based score. We conduct extensive experiments on ImageNet OOD detection benchmarks under diverse settings, including a scenario where the ID data undergoes natural distribution shift. Our results demonstrate that NNGuide provides a significant performance improvement on the base detection scores, achieving state-of-the-art results on both AUROC, FPR95, and AUPR metrics.

We present a novel method called NeTO, for capturing the 3D geometry of solid transparent objects from 2D images via volume rendering. Reconstructing transparent objects is a very challenging task, which is ill-suited for general-purpose reconstruction techniques due to the specular light transport phenomena. Although existing refraction-tracing-based methods, designed especially for this task, achieve impressive results, they still suffer from unstable optimization and loss of fine details since the explicit surface representation they adopted is difficult to be optimized, and the self-occlusion problem is ignored for refraction-tracing. In this paper, we propose to leverage implicit Signed Distance Function (SDF) as surface representation and optimize the SDF field via volume rendering with a self-occlusion aware refractive ray tracing. The implicit representation enables our method to be capable of reconstructing high-quality reconstruction even with a limited set of views, and the self-occlusion aware strategy makes it possible for our method to accurately reconstruct the self-occluded regions. Experiments show that our method achieves faithful reconstruction results and outperforms prior works by a large margin. Visit our project page at https://www.xxlong.site/NeTO/.

Learning a feature point detector presents a challenge both due to the ambiguity of the definition of a keypoint and, correspondingly, the need for specially prepared ground truth labels for such points. In our work, we address both of these issues by utilizing a combination of a hand-crafted Shi-Tomasi detector, a specially designed metric that assesses the quality of keypoints, the stability score (SS), and a neural network. We build on the principled and localized keypoints provided by the Shi-Tomasi detector and learn the neural network to select good feature points via the stability score. The neural network incorporates the knowledge from the training targets in the form of the neural stability score (NeSS). Therefore, our method is named NeSS-ST since it combines the Shi-Tomasi detector and the properties of the neural stability score. It only requires sets of images for training without dataset pre-labeling or the need for reconstructed correspondence labels. We evaluate NeSS-ST on HPatches, ScanNet, MegaDepth and IMC-PT demonstrating state-of-the-art performance and good generalization on downstream tasks. The project repository is available at: https://github.com/KonstantinPakulev/NeSS-ST.

Neural Radiance Fields (NeRF) is a popular neural expression for novel view synthesis. By querying spatial points and view directions, a multilayer perceptron (MLP) can be trained to output the volume density and radiance at each point, which lets us render novel views of the scene. The original NeRF and its recent variants, however, target opaque scenes dominated by diffuse reflection surfaces and cannot handle complex refractive surfaces well. We introduce NeRFrac to realize neural novel view synthesis of scenes captured through refractive surfaces, typically water surfaces. For each queried ray, an MLP-based Refractive Field is trained to estimate the distance from the ray origin to the refractive surface. A refracted ray at each intersection point is then computed by Snell's Law, given the input ray and the approximated local normal. Points of the scene are sampled along the refracted ray and are sent to a Radiance Field for further radiance estimation. We show that from a sparse set of images, our model achieves accurate novel view synthesis of the scene underneath the refractive surface and simultaneously reconstructs the refractive surface. We evaluate the effectiveness of our method with synthetic and real scenes seen through water surfaces. Experimental results demonstrate the accuracy of NeRFrac for modeling scenes seen through wavy refractive surfaces.

Neural radiance fields (NeRF) achieve impressive performance in novel view synthesis when trained on only single sequence data. However, leveraging multiple sequences captured by different cameras at different times is essential for better reconstruction performance. Multi-sequence data takes two main challenges: appearance variation due to different lighting conditions and non-static objects like pedestrians. To address these issues, we propose NeRF-MS, a novel approach to training NeRF with multi-sequence data. Specifically, we utilize a triplet loss to regularize the distribution of per-image appearance code, which leads to better high-frequency texture and consistent appearance, such as specular reflections. Then, we explicitly model non-static objects to reduce floaters. Extensive results demonstrate that NeRF-MS not only outperforms state-of-the-art view synthesis methods on outdoor and synthetic scenes, but also achieves 3D consistent rendering and robust appearance controlling. Project page: https://nerf-ms.github.io/.

Simultaneously odometry and mapping using LiDAR data is an important task for mobile systems to achieve full autonomy in large-scale environments. However, most existing LiDAR-based methods prioritize tracking quality over reconstruction quality. Although the recently developed neural radiance fields (NeRF) have shown promising advances in implicit reconstruction for indoor environments, the problem of simultaneous odometry and mapping for large-scale scenarios using incremental LiDAR data remains unexplored. To bridge this gap, in this paper, we propose a novel NeRF-based LiDAR odometry and mapping approach, NeRF-LOAM, consisting of three modules neural odometry, neural mapping, and mesh reconstruction. All these modules utilize our proposed neural signed distance function, which separates LiDAR points into ground and non-ground points to reduce Z-axis drift, optimizes odometry and voxel embeddings concurrently, and in the end generates dense smooth mesh maps of the environment. Moreover, this joint optimization allows our NeRF-LOAM to be pre-trained free and exhibit strong generalization abilities when applied to different environments. Extensive evaluations on three publicly available datasets demonstrate that our approach achieves state-of-the-art odometry and mapping performance, as well as a strong generalization in large-scale environments utilizing LiDAR data. Furthermore, we perform multiple ablation studies to validate the effectiveness of our network design. The implementation of our approach will be made available at https://github.com/JunyuanDeng/NeRF-LOAM.

We present NeRF-Det, a novel method for indoor 3D detection with posed RGB images as input. Unlike existing indoor 3D detection methods that struggle to model scene geometry, our method makes novel use of NeRF in an end-to-end manner to explicitly estimate 3D geometry, thereby improving 3D detection performance. Specifically, to avoid the significant extra latency associated with per-scene optimization of NeRF, we introduce sufficient geometry priors to enhance the generalizability of NeRF-MLP. Furthermore, we subtly connect the detection and NeRF branches through a shared MLP, enabling an efficient adaptation of NeRF to detection and yielding geometry-aware volumetric representations for 3D detection. Our method outperforms state-of-the-arts by 3.9 mAP and 3.1 mAP on the ScanNet and ARKITScenes benchmarks, respectively. We provide extensive analysis to shed light on how NeRF-Det works. As a result of our joint-training design, NeRF-Det is able to generalize well to unseen scenes for object detection, view synthesis, and depth estimation tasks without requiring per-scene optimization. Code is available at https://github.com/facebookresearch/NeRF-Det.

Recent implicit neural representations have shown great results for novel view synthesis. However, existing methods require expensive per-scene optimization from many views hence limiting their application to real-world unbounded urban settings where the objects of interest or backgrounds are observed from very few views. To mitigate this challenge, we introduce a new approach called NeO 360, Neural fields for sparse view synthesis of outdoor scenes. NeO 360 is a generalizable method that reconstructs 360deg scenes from a single or a few posed RGB images. The essence of our approach is in capturing the distribution of complex real-world outdoor 3D scenes and using a hybrid image-conditional triplanar representation that can be queried from any world point. Our representation combines the best of both voxel-based and bird's-eye-view (BEV) representations and is more effective and expressive than each. NeO 360's representation allows us to learn from a large collection of unbounded 3D scenes while offering generalizability to new views and novel scenes from as few as a single image during inference. We demonstrate our approach on the proposed challenging 360deg unbounded dataset, called NeRDS 360, and show that NeO 360 outperforms state-of-the-art generalizable methods for novel view synthesis while also offering editing and composition capabilities. Project page: zubair-irshad.github.io/projects/neo360.html

Recovering the physical attributes of an object's appearance from its images captured under an unknown illumination is challenging yet essential for photo-realistic rendering.Recent approaches adopt the emerging implicit scene representations and have shown impressive results.However, they unanimously adopt a surface-based representation,and hence can not well handle scenes with very complex geometry, translucent object and etc.In this paper, we propose to conduct inverse volume rendering, in contrast to surface-based, by representing a scene using microflake volume, which assumes the space is filled with infinite small flakes and light reflects or scatters at each spatial location according to microflake distributions. We further adopt the coordinate networks to implicitly encode the microflake volume, and develop a differentiable microflake volume renderer to train the network in an end-to-end way in principle.Our NeMF enables effective recovery of appearance attributes for highly complex geometry and scattering object, enables high-quality relighting, material editing, and especially simulates volume rendering effects, such as scattering, which is infeasible for surface-based approaches. Our data and code are available at: https://github.com/YoujiaZhang/NeMF.

We present a novel differentiable rendering framework for joint geometry, material, and lighting estimation from multi-view images. In contrast to previous methods which assume a simplified environment map or co-located flashlights, in this work, we formulate the lighting of a static scene as one neural incident light field (NeILF) and one outgoing neural radiance field (NeRF). The key insight of the proposed method is the union of the incident and outgoing light fields through physically-based rendering and inter-reflections between surfaces, making it possible to disentangle the scene geometry, material, and lighting from image observations in a physically-based manner. The proposed incident light and inter-reflection framework can be easily applied to other NeRF systems. We show that our method can not only decompose the outgoing radiance into incident lights and surface materials, but also serve as a surface refinement module that further improves the reconstruction detail of the neural surface. We demonstrate on several datasets that the proposed method is able to achieve state-of-the-art results in terms of the geometry reconstruction quality, material estimation accuracy, and the fidelity of novel view rendering.

Images are a convenient way to specify which particular object instance an embodied agent should navigate to. Solving this task requires semantic visual reasoning and exploration of unknown environments. We present a system that can perform this task in both simulation and the real world. Our modular method solves sub-tasks of exploration, goal instance re-identification, goal localization, and local navigation. We re-identify the goal instance in egocentric vision using feature-matching and localize the goal instance by projecting matched features to a map. Each sub-task is solved using off-the-shelf components requiring zero fine-tuning. On the HM3D InstanceImageNav benchmark, this system outperforms a baseline end-to-end RL policy 7x and outperforms a state-of-the-art ImageNav model 2.3x (56% vs. 25% success). We deploy this system to a mobile robot platform and demonstrate effective performance in the real world, achieving an 88% success rate across a home and an office environment.

3D representation disentanglement aims to identify, decompose, and manipulate the underlying explanatory factors of 3D data, which helps AI fundamentally understand our 3D world. This task is currently under-explored and poses great challenges: (i) the 3D representations are complex and in general contains much more information than 2D image; (ii) many 3D representations are not well suited for gradient-based optimization, let alone disentanglement. To address these challenges, we use NeRF as a differentiable 3D representation, and introduce a self-supervised Navigation to identify interpretable semantic directions in the latent space. To our best knowledge, this novel method, dubbed NaviNeRF, is the first work to achieve fine-grained 3D disentanglement without any priors or supervision. Specifically, NaviNeRF is built upon the generative NeRF pipeline, and equipped with an Outer Navigation Branch and an Inner Refinement Branch. They are complementary ---- the outer navigation is to identify global-view semantic directions, and the inner refinement dedicates to fine-grained attributes. A synergistic loss is further devised to coordinate two branches. Extensive experiments demonstrate that NaviNeRF has a superior fine-grained 3D disentanglement ability than the previous 3D-aware models. Its performance is also comparable to editing-oriented models relying on semantic or geometry priors.

Naturally controllable human-scene interaction (HSI) generation has an important role in various fields, such as VR/AR content creation and human-centered AI. However, existing methods are unnatural and unintuitive in their controllability, which heavily limits their application in practice. Therefore, we focus on a challenging task of naturally and controllably generating realistic and diverse HSIs from textual descriptions. From human cognition, the ideal generative model should correctly reason about spatial relationships and interactive actions. To that end, we propose Narrator, a novel relationship reasoning-based generative approach using a conditional variation autoencoder for naturally controllable generation given a 3D scene and a textual description. Also, we model global and local spatial relationships in a 3D scene and a textual description respectively based on the scene graph, and introduce a part-level action mechanism to represent interactions as atomic body part states. In particular, benefiting from our relationship reasoning, we further propose a simple yet effective multi-human generation strategy, which is the first exploration for controllable multi-human scene interaction generation. Our extensive experiments and perceptual studies show that Narrator can controllably generate diverse interactions and significantly outperform existing works.

The long-standing theory that a colour-naming system evolves under dual pressure of efficient communication and perceptual mechanism is supported by more and more linguistic studies, including analysing four decades of diachronic data from the Nafaanra language. This inspires us to explore whether machine learning could evolve and discover a similar colour-naming system via optimising the communication efficiency represented by high-level recognition performance. Here, we propose a novel colour quantisation transformer, CQFormer, that quantises colour space while maintaining the accuracy of machine recognition on the quantised images. Given an RGB image, Annotation Branch maps it into an index map before generating the quantised image with a colour palette; meanwhile the Palette Branch utilises a key-point detection way to find proper colours in the palette among the whole colour space. By interacting with colour annotation, CQFormer is able to balance both the machine vision accuracy and colour perceptual structure such as distinct and stable colour distribution for discovered colour system. Very interestingly, we even observe the consistent evolution pattern between our artificial colour system and basic colour terms across human languages. Besides, our colour quantisation method also offers an efficient quantisation method that effectively compresses the image storage while maintaining high performance in high-level recognition tasks such as classification and detection. Extensive experiments demonstrate the superior performance of our method with extremely low bit-rate colours, showing potential to integrate into quantisation network to quantities from image to network activation. The source code is available at https://github.com/ryeocthiv/CQFormer.

Obtaining personalized 3D animatable avatars from a monocular camera has several real world applications in gaming, virtual try-on, animation, and VR/XR, etc. However, it is very challenging to model dynamic and fine-grained clothing deformations from such sparse data. Existing methods for modeling 3D humans from depth data have limitations in terms of computational efficiency, mesh coherency, and flexibility in resolution and topology. For instance, reconstructing shapes using implicit functions and extracting explicit meshes per frame is computationally expensive and cannot ensure coherent meshes across frames. Moreover, predicting per-vertex deformations on a pre-designed human template with a discrete surface lacks flexibility in resolution and topology. To overcome these limitations, we propose a novel method 'NSF: Neural Surface Fields' for modeling 3D clothed humans from monocular depth. NSF defines a neural field solely on the base surface which models a continuous and flexible displacement field. NSF can be adapted to the base surface with different resolution and topology without retraining at inference time. Compared to existing approaches, our method eliminates the expensive per-frame surface extraction while maintaining mesh coherency, and is capable of reconstructing meshes with arbitrary resolution without retraining. To foster research in this direction, we release our code in project page at: https://yuxuan-xue.com/nsf.

High-fidelity human 3D models can now be learned directly from videos, typically by combining a template-based surface model with neural representations. However, obtaining a template surface requires expensive multi-view capture systems, laser scans, or strictly controlled conditions. Previous methods avoid using a template but rely on a costly or ill-posed mapping from observation to canonical space. We propose a hybrid point-based representation for animatable humans that does not require an explicit surface model, while being generalizable to novel poses. For a given video, our method automatically produces an explicit set of 3D points representing approximate canonical geometry, and learns an articulated deformation model that produces pose-dependent point transformations. The points serve both as a scaffold for high-frequency neural features and an anchor for efficiently mapping between observation and canonical space. We demonstrate on established benchmarks that our representation overcomes limitations of prior work operating in either canonical or in observation space. Moreover, our automatic point extraction approach enables learning models of human and animal characters alike, matching the performance of the methods using rigged surface templates despite being more general. Project website: https: //lemonatsu.github.io/npc/.

Non-line-of-sight (NLOS) imaging is conducted to infer invisible scenes from indirect light on visible objects. The neural transient field (NeTF) was proposed for representing scenes as neural radiance fields in NLOS scenes. We propose NLOS neural implicit surface (NLOS-NeuS), which extends the NeTF to neural implicit surfaces with a signed distance function (SDF) for reconstructing three-dimensional surfaces in NLOS scenes. We introduce two constraints as loss functions for correctly learning an SDF to avoid non-zero level-set surfaces. We also introduce a lower bound constraint of an SDF based on the geometry of the first-returning photons. The experimental results indicate that these constraints are essential for learning a correct SDF in NLOS scenes. Compared with previous methods with discretized representation, NLOS-NeuS with the neural continuous representation enables us to reconstruct smooth surfaces while preserving fine details in NLOS scenes. To the best of our knowledge, this is the first study on neural implicit surfaces with volume rendering in NLOS scenes. Project page: https://yfujimura. github.io/nlos-neus/

Low-light video enhancement in the visible (VIS) range is important yet technically challenging, and it is likely to become more tractable by introducing near-infrared (NIR) information for assistance, which in turn arouses a new challenge on how to obtain appropriate multispectral data for model training. In this paper, we defend the feasibility and superiority of NIR-assisted low-light video enhancement results by using unpaired 24-hour data for the first time, which significantly eases data collection and improves generalization performance on in-the-wild data. By accounting for different physical characteristics between unpaired daytime and nighttime videos, we first propose to turn daytime NIR & VIS into "nighttime mode". Specifically, we design a heuristic yet physics-inspired relighting algorithm to produce realistic pseudo nighttime NIR, and use a resampling strategy followed by a noiseGAN for nighttime VIS conversion. We further devise a temporal-aware network for video enhancement that extracts and fuses bi-directional temporal streams and is trained using real daytime videos and pseudo nighttime videos. We capture multi-spectral data using a co-axial camera and contribute Fulltime Multi-Spectral Video Dataset (FMSVD), the first dataset including aligned 24-hour NIR & VIS videos. Compared to alternative methods, we achieve significantly improved video quality as well as generalization ability on in-the-wild data in terms of both evaluation metrics and visual judgment. Codes and Data Available: https://github.com/MyNiuuu/NVEU.