CVPR 2024 Papers

Out-of-distribution (OOD) detection has attracted a large amount of attention from the machine learning research community in recent years due to its importance in deployed systems. Most of the previous studies focused on the detection of OOD samples in the multi-class classification task. However OOD detection in the multi-label classification task a more common real-world use case remains an underexplored domain. In this research we propose YolOOD - a method that utilizes concepts from the object detection domain to perform OOD detection in the multi-label classification task. Object detection models have an inherent ability to distinguish between objects of interest (in-distribution data) and irrelevant objects (OOD data) in images that contain multiple objects belonging to different class categories. These abilities allow us to convert a regular object detection model into an image classifier with inherent OOD detection capabilities with just minor changes. We compare our approach to state-of-the-art OOD detection methods and demonstrate YolOOD's ability to outperform these methods on a comprehensive suite of in-distribution and OOD benchmark datasets.

Composed Image Retrieval (CIR) is a task that retrieves images similar to a query based on a provided textual modification. Current techniques rely on supervised learning for CIR models using labeled triplets of the <reference image text target image>. These specific triplets are not as commonly available as simple image-text pairs limiting the widespread use of CIR and its scalability. On the other hand zero-shot CIR can be relatively easily trained with image-caption pairs without considering the image-to-image relation but this approach tends to yield lower accuracy. We propose a new semi-supervised CIR approach where we search for a reference and its related target images in auxiliary data and learn our large language model-based Visual Delta Generator (VDG) to generate text describing the visual difference (i.e. visual delta) between the two. VDG equipped with fluent language knowledge and being model agnostic can generate pseudo triplets to boost the performance of CIR models. Our approach significantly improves the existing supervised learning approaches and achieves state-of-the-art results on the CIR benchmarks.

Annotating gaze is an expensive and time-consuming endeavor requiring costly eye-trackers or complex geometric calibration procedures. Although some eye-based unsupervised gaze representation learning methods have been proposed the quality of gaze representation extracted by these methods degrades severely when the head pose is large. In this paper we present the Multi-View Dual-Encoder (MV-DE) a framework designed to learn gaze representations from unlabeled multi-view face images. Through the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy the MV-DE successfully disentangles gaze from general facial information and derives gaze representations closely tied to the subject's eyeball rotation without gaze label. Experimental results illustrate that the gaze representations learned by the MV-DE can be used in downstream tasks including gaze estimation and redirection. Gaze estimation results indicates that the proposed MV-DE displays notably higher robustness to uncontrolled head movements when compared to state-of-the-art (SOTA) unsupervised learning methods.

Recently model merging techniques have surfaced as a solution to combine multiple single-talent models into a single multi-talent model. However previous endeavors in this field have either necessitated additional training or fine-tuning processes or require that the models possess the same pre-trained initialization. In this work we identify a common drawback in prior works w.r.t. the inconsistency of unit similarity in the weight space and the activation space. To address this inconsistency we propose an innovative model merging framework coined as merging under dual-space constraints (MuDSC). Specifically instead of solely maximizing the objective of a single space we advocate for the exploration of permutation matrices situated in a region with a unified high similarity in the dual space achieved through the linear combination of activation and weight similarity matrices. In order to enhance usability we have also incorporated adaptations for group structure including Multi-Head Attention and Group Normalization. Comprehensive experimental comparisons demonstrate that MuDSC can significantly boost the performance of merged models with various task combinations and architectures. Furthermore the visualization of the merged model within the multi-task loss landscape reveals that MuDSC enables the merged model to reside in the overlapping segment featuring a unified lower loss for each task. Our code is publicly available at https://github.com/zju-vipa/training_free_model_merging.

Image warping a classic task in computer vision aims to use geometric transformations to change the appearance of images. Recent methods learn the resampling kernels for warping through neural networks to estimate missing values in irregular grids which however fail to capture local variations in deformed content and produce images with distortion and less high-frequency details. To address this issue this paper proposes an effective method namely MFR to learn Multi-Frequency Representations from input images for image warping. Specifically we propose a progressive filtering network to learn image representations from different frequency subbands and generate deformable images in a coarse-to-fine manner. Furthermore we employ learnable Gabor wavelet filters to improve the model's capability to learn local spatial-frequency representations. Comprehensive experiments including homography transformation equirectangular to perspective projection and asymmetric image super-resolution demonstrate that the proposed MFR significantly outperforms state-of-the-art image warping methods. Our method also showcases superior generalization to out-of-distribution domains where the generated images are equipped with rich details and less distortion thereby high visual quality. The source code is available at https://github.com/junxiao01/MFR.

The rapid advancement of deep learning models is often attributed to their ability to leverage massive training data. In contrast such privilege has not yet fully benefited 3D deep learning mainly due to the limited availability of large-scale 3D datasets. Merging multiple available data sources and letting them collaboratively train a single model is a potential solution. However due to the large domain gap between 3D point cloud datasets such mixed supervision could adversely affect the model's performance and lead to degenerated performance (i.e. negative transfer) compared to single-dataset training. In view of this challenge we introduce Point Prompt Training (PPT) a novel framework for multi-dataset synergistic learning in the context of 3D representation learning that supports multiple pre-training paradigms. Based on this framework we propose Prompt-driven Normalization which adapts the model to different datasets with domain-specific prompts and Language-guided Categorical Alignment that decently unifies the multiple-dataset label spaces by leveraging the relationship between label text. Extensive experiments verify that PPT can overcome the negative transfer associated with synergistic learning and produce generalizable representations. Notably it achieves state-of-the-art performance on each dataset using a single weight-shared model with supervised multi-dataset training. Moreover when served as a pre-training framework it outperforms other pre-training approaches regarding representation quality and attains remarkable state-of-the-art performance across over ten diverse downstream tasks spanning both indoor and outdoor 3D scenarios.

The problem of calibrating deep neural networks (DNNs) for multi-label learning is considered. It is well-known that DNNs trained by cross-entropy for single-label or one-hot classification are poorly calibrated. Many calibration techniques have been proposed to address the problem. However little attention has been paid to the calibration of multi-label DNNs. In this literature the focus has been on improving labeling accuracy in the face of severe dataset unbalance. This is addressed by the introduction of asymmetric losses which have became very popular. However these losses do not induce well calibrated classifiers. In this work we first provide a theoretical explanation for this poor calibration performance by showing that these loses losses lack the strictly proper property a necessary condition for accurate probability estimation. To overcome this problem we propose a new Strictly Proper Asymmetric (SPA) loss. This is complemented by a Label Pair Regularizer (LPR) that increases the number of calibration constraints introduced per training example. The effectiveness of both contributions is validated by extensive experiments on various multi-label datasets. The resulting training method is shown to significantly decrease the calibration error while maintaining state-of-the-art accuracy.

Modeling and visualizing relationships between tasks or datasets is an important step towards solving various meta-tasks such as dataset discovery multi-tasking and transfer learning. However many relationships such as containment and transferability are naturally asymmetric and current approaches for representation and visualization (e.g. t-SNE) do not readily support this. We propose Task2Box an approach to represent tasks using box embeddings---axis-aligned hyperrectangles in low dimensional spaces---that can capture asymmetric relationships between them through volumetric overlaps. We show that Task2Box accurately predicts unseen hierarchical relationships between nodes in ImageNet and iNaturalist datasets as well as transferability between tasks in the Taskonomy benchmark. We also show that box embeddings estimated from task representations (e.g. CLIP Task2Vec or attribute based) can be used to predict relationships between unseen tasks more accurately than classifiers trained on the same representations as well as handcrafted asymmetric distances (e.g. KL divergence). This suggests that low-dimensional box embeddings can effectively capture these task relationships and have the added advantage of being interpretable. We use the approach to visualize relationships among publicly available image classification datasets on popular dataset hosting platform called Hugging Face.

Part-aware panoptic segmentation (PPS) requires (a) that each foreground object and background region in an image is segmented and classified and (b) that all parts within foreground objects are segmented classified and linked to their parent object. Existing methods approach PPS by separately conducting object-level and part-level segmentation. However their part-level predictions are not linked to individual parent objects. Therefore their learning objective is not aligned with the PPS task objective which harms the PPS performance. To solve this and make more accurate PPS predictions we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS). This method uses a set of shared queries to jointly predict (a) object-level segments and (b) the part-level segments within those same objects. As a result TAPPS learns to predict part-level segments that are linked to individual parent objects aligning the learning objective with the task objective and allowing TAPPS to leverage joint object-part representations. With experiments we show that TAPPS considerably outperforms methods that predict objects and parts separately and achieves new state-of-the-art PPS results.

Deep Neural Networks (DNNs) are widely used for their ability to effectively approximate large classes of functions. This flexibility however makes the strict enforcement of constraints on DNNs a difficult problem. In contexts where it is critical to limit the function space to which certain network components belong such as wavelets employed in Multi-Resolution Analysis (MRA) naive constraints via additional terms in the loss function are inadequate. To address this we introduce a Convolutional Neural Network (CNN) wherein the convolutional filters are strictly constrained to be wavelets. This allows the filters to update to task-optimized wavelets during the training procedure. Our primary contribution lies in the rigorous formulation of these filters via a constrained empirical risk minimization framework thereby providing an exact mechanism to enforce these structural constraints. While our work is grounded in theory we investigate our approach empirically through applications in medical imaging particularly in the task of contour prediction around various organs achieving superior performance compared to baseline methods.

General image fusion aims at integrating important information from multi-source images. However due to the significant cross-task gap the respective fusion mechanism varies considerably in practice resulting in limited performance across subtasks. To handle this problem we propose a novel task-customized mixture of adapters (TC-MoA) for general image fusion adaptively prompting various fusion tasks in a unified model. We borrow the insight from the mixture of experts (MoE) taking the experts as efficient tuning adapters to prompt a pre-trained foundation model. These adapters are shared across different tasks and constrained by mutual information regularization ensuring compatibility with different tasks while complementarity for multi-source images. The task-specific routing networks customize these adapters to extract task-specific information from different sources with dynamic dominant intensity performing adaptive visual feature prompt fusion. Notably our TC-MoA controls the dominant intensity bias for different fusion tasks successfully unifying multiple fusion tasks in a single model. Extensive experiments show that TC-MoA outperforms the competing approaches in learning commonalities while retaining compatibility for general image fusion (multi-modal multi-exposure and multi-focus) and also demonstrating striking controllability on more generalization experiments. The code is available at https://github.com/YangSun22/TC-MoA.

Successfully addressing a wide variety of tasks is a core ability of autonomous agents requiring flexibly adapting the underlying decision-making strategies and as we argue in this work also adapting the perception modules. An analogical argument would be the human visual system which uses top-down signals to focus attention determined by the current task. Similarly we adapt pre-trained large vision models conditioned on specific downstream tasks in the context of multi-task policy learning. We introduce task-conditioned adapters that do not require finetuning any pre-trained weights combined with a single policy trained with behavior cloning and capable of addressing multiple tasks. We condition the visual adapters on task embeddings which can be selected at inference if the task is known or alternatively inferred from a set of example demonstrations. To this end we propose a new optimization-based estimator. We evaluate the method on a wide variety of tasks from the CortexBench benchmark and show that compared to existing work it can be addressed with a single policy. In particular we demonstrate that adapting visual features is a key design choice and that the method generalizes to unseen tasks given a few demonstrations.

Exemplar-free Class Incremental Learning (EFCIL) aims to sequentially learn tasks with access only to data from the current one. EFCIL is of interest because it mitigates concerns about privacy and long-term storage of data while at the same time alleviating the problem of catastrophic forgetting in incremental learning. In this work we introduce task-adaptive saliency for EFCIL and propose a new framework which we call Task-Adaptive Saliency Supervision (TASS) for mitigating the negative effects of saliency drift between different tasks. We first apply boundary-guided saliency to maintain task adaptivity and plasticity on model attention. Besides we introduce task-agnostic low-level signals as auxiliary supervision to increase the stability of model attention. Finally we introduce a module for injecting and recovering saliency noise to increase the robustness of saliency preservation. Our experiments demonstrate that our method can better preserve saliency maps across tasks and achieve state-of-the-art results on the CIFAR-100 Tiny-ImageNet and ImageNet-Subset EFCIL benchmarks. Code is available at https://github.com/scok30/tass.

Parkinson's disease (PD) is a devastating movement disorder accelerating in global prevalence but a lack of precision symptom measurement has made the development of effective therapies challenging. The Unified Parkinson's Disease Rating Scale (UPDRS) is the gold-standard for assessing motor symptom severity yet its manual scoring criteria are vague and subjective resulting in coarse and noisy clinical assessments. Machine learning approaches have the potential to modernize PD symptom assessments by making them more quantitative objective and scalable. However the lack of benchmark video datasets for PD motor exams hinders model development. Here we introduce the TULIP dataset to bridge this gap. TULIP emphasizes precision and comprehensiveness comprising multi-view video recordings (6 cameras) of all 25 UPDRS motor exam components together with ratings by 3 clinical experts in a cohort of Parkinson's patients and healthy controls. The multi-view recordings enable 3D reconstructions of body movement that better capture disease signatures than more conventional 2D methods. Using the dataset we establish a baseline model for predicting UPDRS scores from 3D poses illustrating how existing diagnostics could be automated. Looking ahead TULIP could aid the development of new precision diagnostics that transcend UPDRS scores providing a deeper understanding of PD and its potential treatments.

The limited scale of current 3D shape datasets hinders the advancements in 3D shape understanding and motivates multi-modal learning approaches which transfer learned knowledge from data-abundant 2D image and language modalities to 3D shapes. However even though the image and language representations have been aligned by cross-modal models like CLIP we find that the image modality fails to contribute as much as the language in existing multi-modal 3D representation learning methods. This is attributed to the domain shift in the 2D images and the distinct focus of each modality. To more effectively leverage both modalities in the pre-training we introduce TriAdapter Multi-Modal Learning (TAMM) - a novel two-stage learning approach based on three synergistic adapters. First our CLIP Image Adapter mitigates the domain gap between 3D-rendered images and natural images by adapting the visual representations of CLIP for synthetic image-text pairs. Subsequently our Dual Adapters decouple the 3D shape representation space into two complementary sub-spaces: one focusing on visual attributes and the other for semantic understanding which ensure a more comprehensive and effective multi-modal pre-training. Extensive experiments demonstrate that TAMM consistently enhances 3D representations for a wide range of 3D encoder architectures pre-training datasets and downstream tasks. Notably we boost the zero-shot classification accuracy on Objaverse-LVIS from 46.8% to 50.7% and improve the 5-way 10-shot linear probing classification accuracy on ModelNet40 from 96.1% to 99.0%. Project page: https://alanzhangcs.github.io/tamm-page.

Predicting the future occupancy states of the surrounding environment is a vital task for autonomous driving. However current best-performing single-modality methods or multi-modality fusion perception methods are only able to predict uniform snapshots of future occupancy states and require strictly synchronized sensory data for sensor fusion. We propose a novel framework StreamingFlow to lift these strong limitations. StreamingFlow is a novel BEV occupancy predictor that ingests asynchronous multi-sensor data streams for fusion and performs streaming forecasting of the future occupancy map at any future timestamps. By integrating neural ordinary differential equations (N-ODE) into recurrent neural networks StreamingFlow learns derivatives of BEV features over temporal horizons updates the implicit sensor's BEV features as part of the fusion process and propagates BEV states to the desired future time point. It shows good zero-shot generalization ability of prediction reflected in the interpolation of the observed prediction time horizon and the reasonable inference of the unseen farther future period. Extensive experiments on two large-scale datasets nuScenes and Lyft L5 demonstrate that StreamingFlow significantly outperforms previous vision-based LiDAR-based methods and shows superior performance compared to state-of-the-art fusion-based methods.

Prior studies on Remote Sensing Foundation Model (RSFM) reveal immense potential towards a generic model for Earth Observation. Nevertheless these works primarily focus on a single modality without temporal and geo-context modeling hampering their capabilities for diverse tasks. In this study we present SkySense a generic billion-scale model pre-trained on a curated multi-modal Remote Sensing Imagery (RSI) dataset with 21.5 million temporal sequences. SkySense incorporates a factorized multi-modal spatiotemporal encoder taking temporal sequences of optical and Synthetic Aperture Radar (SAR) data as input. This encoder is pre-trained by our proposed Multi-Granularity Contrastive Learning to learn representations across different modal and spatial granularities. To further enhance the RSI representations by the geo-context clue we introduce Geo-Context Prototype Learning to learn region-aware prototypes upon RSI's multi-modal spatiotemporal features. To our best knowledge SkySense is the largest Multi-Modal RSFM to date whose modules can be flexibly combined or used individually to accommodate various tasks. It demonstrates remarkable generalization capabilities on a thorough evaluation encompassing 16 datasets over 7 tasks from single- to multi-modal static to temporal and classification to localization. SkySense surpasses 18 recent RSFMs in all test scenarios. Specifically it outperforms the latest models such as GFM SatLas and Scale-MAE by a large margin i.e. 2.76% 3.67% and 3.61% on average respectively. We will release the pre-trained weights to facilitate future research and Earth Observation applications.

Diffusion models have demonstrated strong potential for robotic trajectory planning. However generating coherent trajectories from high-level instructions remains challenging especially for long-range composition tasks requiring multiple sequential skills. We propose SkillDiffuser an end-to-end hierarchical planning framework integrating interpretable skill learning with conditional diffusion planning to address this problem. At the higher level the skill abstraction module learns discrete human-understandable skill representations from visual observations and language instructions. These learned skill embeddings are then used to condition the diffusion model to generate customized latent trajectories aligned with the skills. This allows generating diverse state trajectories that adhere to the learnable skills. By integrating skill learning with conditional trajectory generation SkillDiffuser produces coherent behavior following abstract instructions across diverse tasks. Experiments on multi-task robotic manipulation benchmarks like Meta-World and LOReL demonstrate state-of-the-art performance and human-interpretable skill representations from SkillDiffuser. More visualization results and information could be found on https://skilldiffuser.github.io/.

We propose SketchINR to advance the representation of vector sketches with implicit neural models. A variable length vector sketch is compressed into a latent space of fixed dimension that implicitly encodes the underlying shape as a function of time and strokes. The learned function predicts the xy point coordinates in a sketch at each time and stroke. Despite its simplicity SketchINR outperforms existing representations at multiple tasks: (i) Encoding an entire sketch dataset into a fixed size latent vector SketchINR gives 60x and 10x data compression over raster and vector sketches respectively. (ii) SketchINR's auto-decoder provides a much higher-fidelity representation than other learned vector sketch representations and is uniquely able to scale to complex vector sketches such as FS-COCO. (iii) SketchINR supports parallelisation that can decode/render 100x faster than other learned vector representations such as SketchRNN. (iv) SketchINR for the first time emulates the human ability to reproduce a sketch with varying abstraction in terms of number and complexity of strokes. As a first look at implicit sketches SketchINR's compact high-fidelity representation will support future work in modelling long and complex sketches.