ICCV 2023 论文列表

We present mSplit, a dedicated approach for trained image decomposition in the context of fluorescence microscopy images. We find that best results using regular deep architectures are achieved when large image patches are used during training, making memory consumption the limiting factor to further improving performance. We therefore introduce lateral contextualization (LC), a novel meta-architecture that enables the memory efficient incorporation of large image-context, which we observe is a key ingredient to solving the image decomposition task at hand. We integrate LC with U-Nets, Hierarchical AEs, and Hierarchical VAEs, for which we formulate a modified ELBO loss. Additionally, LC enables training deeper hierarchical models than otherwise possible and, interestingly, helps to reduce tiling artefacts that are inherently impossible to avoid when using tiled VAE predictions. We apply mSplit to five decomposition tasks, one on a synthetic dataset, four others derived from real microscopy data. Our method consistently achieves best results (average improvements to the best baseline of 2.25 dB PSNR), while simultaneously requiring considerably less GPU memory. Our code and datasets can be found at https://github.com/juglab/uSplit.

Meta-learning-based few-shot detectors use one K-average-pooled prototype (averaging along K-shot dimension) in both Region Proposal Network (RPN) and Detection head (DH) for query detection. Such plain operation would harm the FSOD performance in two aspects: 1) the poor quality of the prototype, and 2) the equivocal guidance due to the contradictions between RPN and DH. In this paper, we look closely into those critical issues and propose the s-Adaptive Decoupled Prototype (s-ADP) as a solution. To generate the high-quality prototype, we prioritize salient representations and deemphasize trivial variations by accessing both angle distance and magnitude dispersion (s) across K-support samples. To provide precise information for the query image, the prototype is decoupled into task-specific ones, which provide tailored guidance for 'where to look' and 'what to look for', respectively. Beyond that, we find our s-ADP can gradually strengthen the generalization power of encoding network during meta-training. So it can robustly deal with intra-class variations and a simple K- average pooling is enough to generate a high-quality prototype at meta-testing. We provide theoretical analysis to support its rationality. Extensive experiments on Pascal VOC, MS-COCO and FSOD datasets demonstrate that the proposed method achieves new state-of-the-art performance. Notably, our method surpasses the baseline model by a large margin - up to around 5.0% AP50 and 8.0% AP75 on novel classes.

Recent advances in implicit neural representations make it possible to generate free-viewpoint videos of the human from sparse view images. To avoid the expensive training for each person, previous methods adopt the generalizable human model and demonstrate impressive results. However, these methods usually rely on limited multi-view images typically collected in the studio or commercial high-quality 3D scans for training, which heavily prohibits their generalization capability for in-the-wild images. To solve this problem, we propose a new approach to learn a generalizable human model from a new source of data, i.e., Internet videos. These videos capture various human appearances and poses and record the performers from abundant viewpoints. To exploit these videos, we present a temporal self-supervised pipeline to enforce the local appearance consistency of each body part over different frames of the same video. Once learned, the human model enables creating photorealistic free-viewpoint videos from a single input image. Experiments show that our method can generate high-quality view synthesis on in-the-wild images while only training on monocular videos.

Existing machine learning (ML) models are often fragile in open environments because the data distribution frequently shifts. To address this problem, domain generalization (DG) aims to explore underlying invariant patterns for stable prediction across domains. In this work, we first characterize that this failure of conventional ML models in DG is attributed to an inadequate identification of causal structures. We further propose a novel and theoretically grounded invariant Directed Acyclic Graph (dubbed iDAG) searching framework that attains an invariant graphical relation as the proxy to the causality structure from the intrinsic data-generating process. To enable tractable computation, iDAG solves a constrained optimization objective built on a set of representative class-conditional prototypes. Additionally, we integrate a hierarchical contrastive learning module, which poses a strong effect of clustering, for enhanced prototypes as well as stabler prediction. Extensive experiments on the synthetic and real-world benchmarks demonstrate that iDAG outperforms the state-of-the-art approaches, verifying the superiority of causal structure identification for DG. The code of iDAG is available at https://github.com/lccurious/iDAG.

Large Language Models (LLMs) have so far impressed the world, with unprecedented capabilities that emerge in models at large scales. On the vision side, transformer models (i.e., ViT) are following the same trend, achieving the best performance on challenging benchmarks. With the abundance of such unimodal models, a natural question arises; do we need also to follow this trend to tackle multimodal tasks? In this work, we propose to rather direct effort to efficient adaptations of existing models, and propose to augment Language Models with perception. Existing approaches for adapting pretrained models for vision-language tasks still rely on several key components that hinder their efficiency. In particular, they still train a large number of parameters, rely on large multimodal pretraining, use encoders (e.g., CLIP) trained on huge image-text datasets, and add significant inference overhead. In addition, most of these approaches have focused on Zero-Shot and In Context Learning, with little to no effort on direct finetuning. We investigate the minimal computational effort needed to adapt unimodal models for multimodal tasks and propose a new challenging setup, alongside different approaches, that efficiently adapts unimodal pretrained models. We show that by freezing more than 99% of total parameters, training only one linear projection layer, and prepending only one trainable token, our approach (dubbed eP-ALM) significantly outperforms other baselines on VQA and Captioning across Image, Video, and Audio modalities, following the proposed setup. The code is available here: https://github.com/mshukor/eP-ALM.

As it is hard to calibrate single-view RGB images in the wild, existing 3D human mesh reconstruction (3DHMR) methods either use a constant large focal length or estimate one based on the background environment context, which can not tackle the problem of the torso, limb, hand or face distortion caused by perspective camera projection when the camera is close to the human body. The naive focal length assumptions can harm this task with the incorrectly formulated projection matrices. To solve this, we propose Zolly, the first 3DHMR method focusing on perspective-distorted images. Our approach begins with analysing the reason for perspective distortion, which we find is mainly caused by the relative location of the human body to the camera center. We propose a new camera model and a novel 2D representation, termed distortion image, which describes the 2D dense distortion scale of the human body. We then estimate the distance from distortion scale features rather than environment context features. Afterwards, We integrate the distortion feature with image features to reconstruct the body mesh. To formulate the correct projection matrix and locate the human body position, we simultaneously use perspective and weak-perspective projection loss. Since existing datasets could not handle this task, we propose the first synthetic dataset PDHuman and extend two real-world datasets tailored for this task, all containing perspective-distorted human images. Extensive experiments show that Zolly outperforms existing state-of-the-art methods on both perspective-distorted datasets and the standard benchmark (3DPW). Code and dataset will be released at https://wenjiawang0312.github.io/projects/zolly/.

Neural Radiance Field training can be accelerated through the use of grid-based representations in NeRF's learned mapping from spatial coordinates to colors and volumetric density. However, these grid-based approaches lack an explicit understanding of scale and therefore often introduce aliasing, usually in the form of jaggies or missing scene content. Anti-aliasing has previously been addressed by mip-NeRF 360, which reasons about sub-volumes along a cone rather than points along a ray, but this approach is not natively compatible with current grid-based techniques. We show how ideas from rendering and signal processing can be used to construct a technique that combines mip-NeRF 360 and grid-based models such as Instant NGP to yield error rates that are 8%-77% lower than either prior technique, and that trains 24x faster than mip-NeRF 360.

The joint visual-language model CLIP has enabled new and exciting applications, such as open-vocabulary segmentation, which can locate any segment given an arbitrary text query. In our research, we ask whether it is possible to discover semantic segments without any user guidance in the form of text queries or predefined classes, and label them using natural language automatically? We propose a novel problem zero-guidance segmentation and the first baseline that leverages two pre-trained generalist models, DINO and CLIP, to solve this problem without any fine-tuning or segmentation dataset. The general idea is to first segment an image into small over-segments, encode them into CLIP's visual-language space, translate them into text labels, and merge semantically similar segments together. The key challenge, however, is how to encode a visual segment into a segment-specific embedding that balances global and local context information, both useful for recognition. Our main contribution is a novel attention-masking technique that balances the two contexts by analyzing the attention layers inside CLIP. We also introduce several metrics for the evaluation of this new task. With CLIP's innate knowledge, our method can precisely locate the Mona Lisa painting among a museum crowd.

Large-scale text-to-image diffusion models have significantly improved the state of the art in generative image modeling and allow for an intuitive and powerful user interface to drive the image generation process. Expressing spatial constraints, e.g. to position specific objects in particular locations, is cumbersome using text; and current text-based image generation models are not able to accurately follow such instructions. In this paper we consider image generation from text associated with segments on the image canvas, which combines an intuitive natural language interface with precise spatial control over the generated content. We propose ZestGuide, a "ZEro-shot" SegmenTation Guidance approach that can be plugged into pre-trained text-to-image diffusion models, and does not require any additional training. It leverages implicit segmentation maps that can be extracted from cross-attention layers, and uses them to align the generation with input masks. Our experimental results combine high image quality with accurate alignment of generated content with input segmentations, and improve over prior work both quantitatively and qualitatively, including methods that require training on images with corresponding segmentations. Compared to Paint with Words, the previous state-of-the art in image generation with zero-shot segmentation conditioning, we improve by 5 to 10 mIoU points on the COCO dataset with similar FID scores.

This paper proposes a feature synthesis approach for zero-shot semantic segmentation of 3D point clouds, enabling generalization to previously unseen categories. Given only the class-level semantic information for unseen objects, we strive to enhance the correspondence, alignment and consistency between the visual and semantic spaces, to synthesise diverse, generic and transferable visual features. We develop a masked learning strategy to promote diversity within the same class visual features and enhance the separation between different classes. We further cast the visual features into a prototypical space to model their distribution for alignment with the corresponding semantic space. Finally, we develop a consistency regularizer to preserve the semantic-visual relationships between the real-seen features and synthetic-unseen features. Our approach shows considerable semantic segmentation gains on ScanNet, S3DIS and SemanticKITTI benchmarks. Our code is available at: https://github.com/leolyj/3DPC-GZSL

Diffusion models have shown great promise in text-guided image style transfer, but there is a trade-off between style transformation and content preservation due to their stochastic nature. Existing methods require computationally expensive fine-tuning of diffusion models or additional neural network. To address this, here we propose a zero-shot contrastive loss for diffusion models that doesn't require additional fine-tuning or auxiliary networks. By leveraging patch-wise contrastive loss between generated samples and original image embeddings in the pre-trained diffusion model, our method can generate images with the same semantic content as the source image in a zero-shot manner. Our approach outperforms existing methods while preserving content and requiring no additional training, not only for image style transfer but also for image-to-image translation and manipulation. Our experimental results validate the effectiveness of our proposed method.

Composed Image Retrieval (CIR) aims to retrieve a target image based on a query composed of a reference image and a relative caption that describes the difference between the two images. The high effort and cost required for labeling datasets for CIR hamper the widespread usage of existing methods, as they rely on supervised learning. In this work, we propose a new task, Zero-Shot CIR (ZS-CIR), that aims to address CIR without requiring a labeled training dataset. Our approach, named zero-Shot composEd imAge Retrieval with textuaL invErsion SEARLE, maps the visual features of the reference image into a pseudo-word token in CLIP token embedding space and integrates it with the relative caption. To support research on ZS-CIR, we introduce an open-domain benchmarking dataset named Composed Image Retrieval on Common Objects in context (CIRCO), which is the first dataset for CIR containing multiple ground-truths for each query. The experiments show that SEARLE exhibits better performance than the baselines on the two main datasets for CIR tasks, FashionIQ and CIRR, and on the proposed CIRCO. The dataset, the code and the model are publicly available at https://github.com/miccunifi/SEARLE.

We introduce Zero-1-to-3, a framework for changing the camera viewpoint of an object given just a single RGB image. To perform novel view synthesis in this underconstrained setting, we capitalize on the geometric priors that large-scale diffusion models learn about natural images. Our conditional diffusion model uses a synthetic dataset to learn controls of the relative camera viewpoint, which allow new images to be generated of the same object under a specified camera transformation. Even though it is trained on a synthetic dataset, our model retains a strong zero-shot generalization ability to out-of-distribution datasets as well as in-the-wild images, including impressionist paintings. Our viewpoint-conditioned diffusion approach can further be used for the task of 3D reconstruction from a single image. Qualitative and quantitative experiments show that our method significantly outperforms stateof- the-art single-view 3D reconstruction and novel view synthesis models by leveraging Internet-scale pre-training.

Existing datasets for autonomous driving (AD) often lack diversity and long-range capabilities, focusing instead on 360* perception and temporal reasoning. To address this gap, we introduce ZOD, a large-scale and diverse multimodal dataset collected over two years in various European countries, covering an area 9x that of existing datasets. ZOD boasts the highest range and resolution sensors among comparable datasets, coupled with detailed keyframe annotations for 2D and 3D objects (up to 245m), road instance/semantic segmentation, traffic sign recognition, and road classification. We believe that this unique combination will facilitate breakthroughs in long-range perception and multi-task learning. The dataset is composed of Frames, Sequences, and Drives, designed to encompass both data diversity and support for spatio-temporal learning, sensor fusion, localization, and mapping. Frames consist of 100k curated camera images with two seconds of other supporting sensor data, while the 1473 Sequences and 29 Drives include the entire sensor suite for 20 seconds and a few minutes, respectively. ZOD is the only AD dataset released under the permissive CC BY-SA 4.0 license, allowing for both research and commercial use. More information, and an extensive devkit, can be found at zod.zenseact.com.

The recent wave of large-scale text-to-image diffusion models has dramatically increased our text-based image generation abilities. These models can generate realistic images for a staggering variety of prompts and exhibit impressive compositional generalization abilities. Almost all use cases thus far have solely focused on sampling; however, diffusion models can also provide conditional density estimates, which are useful for tasks beyond image generation. In this paper, we show that the density estimates from large-scale text-to-image diffusion models like Stable Diffusion can be leveraged to perform zero-shot classification without any additional training. Our generative approach to classification, which we call Diffusion Classifier, attains strong results on a variety of benchmarks and outperforms alternative methods of extracting knowledge from diffusion models. Although a gap remains between generative and discriminative approaches on zero-shot recognition tasks, our diffusion-based approach has stronger multimodal compositional reasoning abilities than competing discriminative approaches. Finally, we use Diffusion Classifier to extract standard classifiers from class-conditional diffusion models trained on ImageNet. These models approach the performance of SOTA discriminative classifiers and exhibit strong "effective robustness" to distribution shift. Overall, our results are a step toward using generative over discriminative models for downstream tasks.

We propose SeedAL, a method to seed active learning for efficient annotation of 3D point clouds for semantic segmentation. Active Learning (AL) iteratively selects relevant data fractions to annotate within a given budget, but requires a first fraction of the dataset (a 'seed') to be already annotated to estimate the benefit of annotating other data fractions. We first show that the choice of the seed can significantly affect the performance of many AL methods. We then propose a method for automatically constructing a seed that will ensure good performance for AL. Assuming that images of the point clouds are available, which is common, our method relies on powerful unsupervised image features to measure the diversity of the point clouds. It selects the point clouds for the seed by optimizing the diversity under an annotation budget, which can be done by solving a linear optimization problem. Our experiments demonstrate the effectiveness of our approach compared to random seeding and existing methods on both the S3DIS and SemanticKitti datasets. Code is available at https://github.com/nerminsamet/seedal.

Large-scale visual localization systems continue to relyon 3D point clouds built from image collections usingstructure-from-motion. While the 3D points in these modelsare represented using local image features, directly match-ing a query image's local features against the point cloud ischallenging due to the scale of the nearest-neighbor searchproblem. Many recent approaches to visual localization havethus proposed a hybrid method, where first a global (per im-age) embedding is used to retrieve a small subset of databaseimages, and local features of the query are matched onlyagainst those. It seems to have become common belief thatglobal embeddings are critical for said image-retrieval invisual localization, despite the significant downside of hav-ing to compute two feature types for each query image. Inthis paper, we take a step back from this assumption and pro-pose Constrained Approximate Nearest Neighbors (CANN),a joint solution of k-nearest-neighbors across both the ge-ometry and appearance space using only local features. Wefirst derive the theoretical foundation for k-nearest-neighborretrieval across multiple metrics and then showcase howCANN improves visual localization. Our experiments onpublic localization benchmarks demonstrate that our methodsignificantly outperforms both state-of-the-art global feature-based retrieval and approaches using local feature aggrega-tion schemes. Moreover, it is an order of magnitude faster inboth index and query time than feature aggregation schemesfor these datasets. Code will be released.

The recent interest in the edge-to-cloud continuum paradigm has emphasized the need for simple and scalable architectures to deliver optimal performance on computationally constrained devices. However, resource-efficient neural networks usually optimize for parameter count and thus use operators such as depthwise convolutions, which do not maximally exploit the efficiency of resource-constrained devices. In this article, we propose XiNet, a novel convolutional neural architecture that targets edge devices. We derived the XiNet architecture from an extensive real-world efficiency analysis of various neural network operators (e.g., standard, depthwise, and pointwise convolutions). Compared to other mobile architectures, our approach substantially improves the performance-complexity trade-off by optimizing the number of operations, parameters, and working memory (RAM). Moreover, we show how XiNet can be easily adapted to different devices thanks to Hardware Aware Scaling (HAS), which enables disjoint optimization of RAM, FLASH, and operations count. We analyze the scaling properties of our architecture under different hardware constraints and validate it on the image classification task. Finally, we evaluate the performance of XiNet for object detection on the MS-COCO and VOC-2012 benchmarks and compare it with state-of-the-art mobile neural networks, achieving a 70% reduction in energy requirements with similar performance.

We propose XVO, a semi-supervised learning method for training generalized monocular Visual Odometry (VO) models with robust off-the-self operation across diverse datasets and settings. In contrast to standard monocular VO approaches which often study a known calibration within a single dataset, XVO efficiently learns to recover relative pose with real-world scale from visual scene semantics, i.e., without relying on any known camera parameters. We optimize the motion estimation model via self-training from large amounts of unconstrained and heterogeneous dash camera videos available on YouTube. Our key contribution is twofold. First, we empirically demonstrate the benefits of semi-supervised training for learning a general-purpose direct VO regression network. Second, we demonstrate multi-modal supervision, including segmentation, flow, depth, and audio auxiliary prediction tasks, to facilitate generalized representations for the VO task. Specifically, we find audio prediction task to significantly enhance the semi-supervised learning process while alleviating noisy pseudo-labels, particularly in highly dynamic and out-of-domain video data. Our proposed teacher network achieves state-of-the-art performance on the commonly used KITTI benchmark despite no multi-frame optimization or knowledge of camera parameters. Combined with the proposed semi-supervised step, XVO demonstrates off-the-shelf knowledge transfer across diverse conditions on KITTI, nuScenes, and Argoverse without fine-tuning.