ICCV 2023 Papers

Multi-modality image fusion and segmentation play a vital role in autonomous driving and robotic operation. Early efforts focus on boosting the performance for only one task, e.g., fusion or segmentation, making it hard to reach `Best of Both Worlds'. To overcome this issue, in this paper, we propose a Multi-interactive Feature learning architecture for image fusion and segmentation, namely SegMiF, and exploit dual-task correlation to promote the performance of both tasks. The SegMiF is of a cascade structure, containing a fusion sub-network and a commonly used segmentation sub-network. By slickly bridging intermediate features between two components, the knowledge learned from the segmentation task can effectively assist the fusion task. Also, the benefited fusion network supports the segmentation one to perform more pretentiously. Besides, a hierarchical interactive attention block is established to ensure fine-grained mapping of all the vital information between two tasks, so that the modality/semantic features can be fully mutual-interactive. In addition, a dynamic weight factor is introduced to automatically adjust the corresponding weights of each task, which can balance the interactive feature correspondence and break through the limitation of laborious tuning. Furthermore, we construct a smart multi-wave binocular imaging system and collect a full-time multi-modality benchmark with 15 annotated pixel-level categories for image fusion and segmentation. Extensive experiments on several public datasets and our benchmark demonstrate that the proposed method outputs visually appealing fused images and perform averagely 7.66% higher segmentation mIoU in the real-world scene than the state-of-the-art approaches. The source code and benchmark are available at https://github.com/JinyuanLiu-CV/SegMiF.

Synthesizing high-fidelity videos from real-world multiview input is challenging due to the complexities of real-world environments and high-dynamic movements. Previous works based on neural radiance fields have demonstrated high-quality reconstructions of dynamic scenes. However, training such models on real-world scenes is time-consuming, usually taking days or weeks. In this paper, we present a novel method named MixVoxels to efficiently represent dynamic scenes which leads to fast training and rendering speed. The proposed MixVoxels represents the 4D dynamic scenes as a mixture of static and dynamic voxels and processes them with different networks. In this way, the computation of the required modalities for static voxels can be processed by a lightweight model, which essentially reduces the amount of computation as many daily dynamic scenes are dominated by static backgrounds. To distinguish the two kinds of voxels, we propose a novel variation field to estimate the temporal variance of each voxel. For the dynamic representations, we design an inner-product time query method to efficiently query multiple time steps, which is essential to recover the high-dynamic movements. As a result, with 15 minutes of training for dynamic scenes with inputs of 300-frame videos, MixVoxels achieves better PSNR than previous methods. For rendering, MixVoxels can render a novel view video with 1K resolution at 37 fps. Codes and trained models are available at https://github.com/fengres/mixvoxels.

In video action recognition, shortcut static features can interfere with the learning of motion features, resulting in poor out-of-distribution (OOD) generalization. The video background is clearly a source of static bias, but the video foreground, such as the clothing of the actor, can also provide static bias. In this paper, we empirically verify the existence of foreground static bias by creating test videos with conflicting signals from the static and moving portions of the video. To tackle this issue, we propose a simple yet effective technique, StillMix, to learn robust action representations. Specifically, StillMix identifies bias-inducing video frames using a 2D reference network and mixes them with videos for training, serving as effective bias suppression even when we cannot explicitly extract the source of bias within each video frame or enumerate types of bias. Finally, to precisely evaluate static bias, we synthesize two new benchmarks, SCUBA for static cues in the background, and SCUFO for static cues in the foreground. With extensive experiments, we demonstrate that StillMix mitigates both types of static bias and improves video representations for downstream applications.

For a generalized (or non-central) camera model, the minimal problem for two views of six points has efficient solvers. However, minimal problems of three views with four points and three views of six lines have not yet been explored and solved, despite the efforts from the computer vision community. This paper develops the formulations of these two minimal problems and shows how state-of-the-art GPU implementations of Homotopy Continuation solver can be used effectively. The proposed methods are evaluated on both synthetic and real datasets, demonstrating that they are fast, accurate and that they improve on structure from motion estimations, when employed in an hypothesis and test setting.

The field of visual representation learning has seen explosive growth in the past years, but its benefits in robotics have been surprisingly limited so far. Prior work uses generic visual representations as a basis to learn (task-specific) robot action policies (e.g., via behavior cloning). While the visual representations do accelerate learning, they are primarily used to encode visual observations. Thus, action information has to be derived purely from robot data, which is expensive to collect! In this work, we present a scalable alternative where the visual representations can help directly infer robot actions. We observe that vision encoders express relationships between image observations as distances (e.g., via embedding dot product) that could be used to efficiently plan robot behavior. We operationalize this insight and develop a simple algorithm for acquiring a distance function and dynamics predictor, by fine-tuning a pre-trained representation on human collected video sequences. The final method is able to substantially outperform traditional robot learning baselines (e.g., 70% success v.s. 50% for behavior cloning on pick-place) on a suite of diverse real-world manipulation tasks. It can also generalize to novel objects, without using any robot demonstrations during train time. For visualizations of the learned policies please check: https://agi-labs.github.io/manipulate-by-seeing/

We tackle the challenging task of unsupervised object localization in this work. Recently, transformers trained with self-supervised learning have been shown to exhibit object localization properties without being trained for this task. In this work, we present Multiple Object localization with Self-supervised Transformers (MOST) that uses features of transformers trained using self-supervised learning to localize multiple objects in real world images. MOST analyzes the similarity maps of the features using box counting; a fractal analysis tool to identify tokens lying on foreground patches. The identified tokens are then clustered together, and tokens of each cluster are used to generate bounding boxes on foreground regions. Unlike recent state-of-the-art object localization methods, MOST can localize multiple objects per image and outperforms SOTA algorithms on several object localization and discovery benchmarks on PASCAL-VOC 07, 12 and COCO20k datasets. Additionally, we show that MOST can be used for self-supervised pretraining of object detectors, and yields consistent improvements on fully, semi-supervised object detection and unsupervised region proposal generation.Our project is publicly available at rssaketh.github.io/most.

A central challenge of video prediction lies where the system has to reason the object's future motion from image frames while simultaneously maintaining the consistency of its appearance across frames. This work introduces an end-to-end trainable two-stream video prediction framework, Motion-Matrix-based Video Prediction (MMVP), to tackle this challenge. Unlike previous methods that usually handle motion prediction and appearance maintenance within the same set of modules, MMVP decouples motion and appearance information by constructing appearance-agnostic motion matrices. The motion matrices represent the temporal similarity of each and every pair of feature patches in the input frames, and are the sole input of the motion prediction module in MMVP. This design improves video prediction in both accuracy and efficiency, and reduces the model size. Results of extensive experiments demonstrate that MMVP outperforms state-of-the-art systems on public data sets by non-negligible large margins (approx. 1 db in PSNR, UCF Sports) in significantly smaller model sizes (84% the size or smaller). Please refer to https://github.com/Kay1794/MMVP-motion-matrix-based-video-prediction for the official code and the datasets used in this paper.

Deep learning has achieved tremendous success in recent years, but most of these successes are built on an independent and identically distributed (IID) assumption. This somewhat hinders the application of deep learning to the more challenging out-of-distribution (OOD) scenarios. Although many OOD methods have been proposed to address this problem and have obtained good performance on testing data that is of major shifts with training distributions, interestingly, we experimentally find that these methods achieve excellent OOD performance by making a great sacrifice of the IID performance. We call this finding the IID-OOD dilemma. Clearly, in real-world applications, distribution shifts between training and testing data are often uncertain, where shifts could be minor, and even close to the IID scenario, and thus it is truly important to design a deep model with the balanced generalization ability between IID and OOD. To this end, in this paper, we investigate an intriguing problem of balancing IID and OOD generalizations and propose a novel Model Agnostic adaPters (MAP) method, which is more reliable and effective for distribution-shift-agnostic real-world data. Our key technical contribution is to use auxiliary adapter layers to incorporate the inductive bias of IID into OOD methods. To achieve this goal, we apply a bilevel optimization to explicitly model and optimize the coupling relationship between the OOD model and auxiliary adapter layers. We also theoretically give a first-order approximation to save computational time. Experimental results on six datasets successfully demonstrate that MAP can greatly improve the performance of IID while achieving good OOD performance.

Current high-performance semantic segmentation models are purely data-driven sub-symbolic approaches and blind to the structured nature of the visual world. This is in stark contrast to human cognition which abstracts visual perceptions at multiple levels and conducts symbolic reasoning with such structured abstraction. To fill these fundamental gaps, we devise LogicSeg, a holistic visual semantic parser that integrates neural inductive learning and logic reasoning with both rich data and symbolic knowledge. In particular, the semantic concepts of interest are structured as a hierarchy, from which a comprehensive set of constraints are derived for describing the symbolic relations and formalized in first-order logic. After fuzzy logic-based continuous relaxation, logical formulae are grounded onto data and neural computational graphs, hence enabling logic-induced network training. During inference, logical constraints are packaged into an iterative process and injected into the network in a form of several matrix multiplications, so as to achieve hierarchy-coherent prediction with logic reasoning. These designs together make LogicSeg a general and compact neural-logic machine that is readily integrated into existing segmentation models. Extensive experiments over four datasets with various segmentation models and backbones verify the effectiveness and generality of LogicSeg. We believe this study opens a new avenue for visual semantic parsing. Our code will be released.

Dense 3D reconstruction from RGB images traditionally assumes static camera pose estimates. This assumption has endured, even as recent works have increasingly focused on real-time methods for mobile devices. However, the assumption of a fixed pose for each image does not hold for online execution: poses from real-time SLAM are dynamic and may be updated following events such as bundle adjustment and loop closure. This has been addressed in the RGB-D setting, by de-integrating past views and re-integrating them with updated poses, but it remains largely untreated in the RGB-only setting. We formalize this problem to define the new task of dense online reconstruction from dynamically-posed images. To support further research, we introduce a dataset called LivePose containing the dynamic poses from a SLAM system running on ScanNet. We select three recent reconstruction systems and apply a framework based on de-integration to adapt each one to the dynamic-pose setting. In addition, we propose a novel, non-linear de-integration module that learns to remove stale scene content. We show that responding to pose updates is critical for high-quality reconstruction, and that our de-integration framework is an effective solution.

We introduce LiDAR-UDA, a novel two-stage self-training-based Unsupervised Domain Adaptation (UDA) method for LiDAR segmentation. Existing self-training methods use a model trained on labeled source data to generate pseudo labels for target data and refine the predictions via fine-tuning the network on the pseudo labels. These methods suffer from domain shifts caused by different LiDAR sensor configurations in the source and target domains. We propose two techniques to reduce sensor discrepancy and improve pseudo label quality: 1) LiDAR beam subsampling, which simulates different LiDAR scanning patterns by randomly dropping beams; 2) cross-frame ensembling, which exploits temporal consistency of consecutive frames to generate more reliable pseudo labels. Our method is simple, generalizable, and does not incur any extra inference cost. We evaluate our method on several public LiDAR datasets and show that it outperforms the state-of-the-art methods by more than 3.9% mIoU on average for all scenarios. Code will be available at https://github.com/JHLee0513/lidar_uda.

Location determination finds wide applications in daily life. Instead of existing efforts devoted to localizing tourist photos captured by perspective cameras, in this article, we focus on developing person positioning solutions using overhead fisheye cameras. Such solutions are advantageous in large field of view (FOV), low cost, anti-occlusion, and unaggressive work mode (without the necessity of cameras carried by persons). However, related studies are quite scarce, due to the paucity of data. To stimulate research in this exciting area, we present LOAF, the first large-scale overhead fisheye dataset for person detection and localization. LOAF is built with many essential features, e.g., i) the data cover abundant diversities in scenes, human pose, density, and location; ii) it contains currently the largest number of annotated pedestrian, i.e., 600K bounding boxes with ground-truth location information; iii) the body-boxes are labeled as radius-aligned so as to fully address the positioning challenge. To approach localization, we build a fisheye person detection network, which exploits the fisheye distortions by a clever position embedding strategy and is trained to predict radius-aligned human boxes end-to-end. Then, the actual locations of the detected persons are calculated by a numerical solution on the fisheye model and camera altitude data. Extensive experiments on LOAF validate the superiority of our fisheye detector w.r.t. previous methods, and show that our whole fisheye positioning solution is able to locate all persons in FOV with an accuracy of 0.5m, within 0.1s. Our dataset and code shall be released.

Recognizing objects from sparse and noisy events becomes extremely difficult when paired images and category labels do not exist. In this paper, we study label-free event-based object recognition where category labels and paired images are not available. To this end, we propose a joint formulation of object recognition and image reconstruction in a complementary manner. Our method first reconstructs images from events and performs object recognition through Contrastive Language-Image Pre-training (CLIP), enabling better recognition through a rich context of images. Since the category information is essential in reconstructing images, we propose category-guided attraction loss and category-agnostic repulsion loss to bridge the textual features of predicted categories and the visual features of reconstructed images using CLIP. Moreover, we introduce a reliable data sampling strategy and local-global reconstruction consistency to boost joint learning of two tasks. To enhance the accuracy of prediction and quality of reconstruction, we also propose a prototype-based approach using unpaired images. Extensive experiments demonstrate the superiority of our method and its extensibility for zero-shot object recognition. Our project code is available at https://github.com/Chohoonhee/Ev-LaFOR.

Humans describe the physical world using natural language to refer to specific 3D locations based on a vast range of properties: visual appearance, semantics, abstract associations, or actionable affordances. In this work we propose Language Embedded Radiance Fields (LERFs), a method for grounding language embeddings from off-the-shelf models like CLIP into NeRF, which enable these types of open-ended language queries in 3D. LERF learns a dense, multi-scale language field inside NeRF by volume rendering CLIP embeddings along training rays, supervising these embeddings across training views to provide multi-view consistency and smooth the underlying language field. After optimization, LERF can extract 3D relevancy maps for a broad range of language prompts interactively in real-time, which has potential use cases in robotics, understanding vision-language models, and interacting with 3D scenes. LERF enables pixel-aligned, zero-shot queries on the distilled 3D CLIP embeddings without relying on region proposals or masks, supporting long-tail open-vocabulary queries hierarchically across the volume. See the project website at: https://lerf.io

Large-scale pre-training tasks like image classification, captioning, or self-supervised techniques do not incentivize learning the semantic boundaries of objects. However, recent generative foundation models built using text-based latent diffusion techniques may learn semantic boundaries. This is because they have to synthesize intricate details about all objects in an image based on a text description. Therefore, we present a technique for segmenting real and AI-generated images using latent diffusion models (LDMs) trained on internet-scale datasets. First, we show that the latent space of LDMs (z-space) is a better input representation compared to other feature representations like RGB images or CLIP encodings for text-based image segmentation. By training the segmentation models on the latent z-space, which creates a compressed representation across several domains like different forms of art, cartoons, illustrations, and photographs, we are also able to bridge the domain gap between real and AI-generated images. We show that the internal features of LDMs contain rich semantic information and present a technique in the form of LD-ZNet to further boost the performance of text-based segmentation. Overall, we show up to 6% improvement over standard baselines for text-to-image segmentation on natural images. For AI-generated imagery, we show close to 20% improvement compared to state-of-the-art techniques. The project is available at https://koutilya-pnvr.github.io/LD-ZNet/.

3D lane detection from monocular images is a fundamental yet challenging task in autonomous driving. Recent advances primarily rely on structural 3D surrogates (e.g., bird's eye view) built from front-view image features and camera parameters. However, the depth ambiguity in monocular images inevitably causes misalignment between the constructed surrogate feature map and the original image, posing a great challenge for accurate lane detection. To address the above issue, we present a novel LATR model, an end-to-end 3D lane detector that uses 3D-aware front-view features without transformed view representation. Specifically, LATR detects 3D lanes via cross-attention based on query and key-value pairs, constructed using our lane-aware query generator and dynamic 3D ground positional embedding. On the one hand, each query is generated based on 2D lane-aware features and adopts a hybrid embedding to enhance the lane information. On the other hand, 3D space information is injected as positional embedding from an iteratively-updated 3D ground plane. LATR outperforms previous state-of-the-art methods on both synthetic Apollo and realistic OpenLane, ONCE-3DLanes datasets by large margins (e.g., 11.4 gain in terms of F1 score on OpenLane). Code will be released at https://github.com/JMoonr/LATR.

We propose a novel unsupervised backlit image enhancement method, abbreviated as CLIP-LIT, by exploring the potential of Contrastive Language-Image Pre-Training (CLIP) for pixel-level image enhancement. We show that the open-world CLIP prior not only aids in distinguishing between backlit and well-lit images, but also in perceiving heterogeneous regions with different luminance, facilitating the optimization of the enhancement network. Unlike high-level and image manipulation tasks, directly applying CLIP to enhancement tasks is non-trivial, owing to the difficulty in finding accurate prompts. To solve this issue, we devise a prompt learning framework that first learns an initial prompt pair by constraining the text-image similarity between the prompt (negative/positive sample) and the corresponding image (backlit image/well-lit image) in the CLIP latent space. Then, we train the enhancement network based on the text-image similarity between the enhanced result and the initial prompt pair. To further improve the accuracy of the initial prompt pair, we iteratively fine-tune the prompt learning framework to reduce the distribution gaps between the backlit images, enhanced results, and well-lit images via rank learning, boosting the enhancement performance. Our method alternates between updating the prompt learning framework and enhancement network until visually pleasing results are achieved. Extensive experiments demonstrate that our method outperforms state-of-the-art methods in terms of visual quality and generalization ability, without requiring any paired data.

In this paper, we propose a novel task IntentQA, a special VideoQA task focusing on video intent reasoning, which has become increasingly important for AI with its advantages in equipping AI agents with the capability of reasoning beyond mere recognition in daily tasks. We also contribute a large-scale VideoQA dataset for this task. We propose a Context-aware Video Intent Reasoning model (CaVIR) consisting of i) Video Query Language (VQL) for better cross-modal representation of the situational context, ii) Contrastive Learning module for utilizing the contrastive context, and iii) Commonsense Reasoning module for incorporating the commonsense context. Comprehensive experiments on this challenging task demonstrate the effectiveness of each model component, the superiority of our full model over other baselines, and the generalizability of our model to a new VideoQA task. The dataset and codes are open-sourced at: https://github.com/JoseponLee/IntentQA.git

We propose a method for editing NeRF scenes with text-instructions. Given a NeRF of a scene and the collection of images used to reconstruct it, our method uses an image-conditioned diffusion model (InstructPix2Pix) to iteratively edit the input images while optimizing the underlying scene, resulting in an optimized 3D scene that respects the edit instruction. We demonstrate that our proposed method is able to edit large-scale, real-world scenes, and is able to accomplish more realistic, targeted edits than prior work.