CVPR 2024 Papers

Vision-language models (VLMs) are trained for thousands of GPU hours on carefully selected subsets of massive web scrapes. For instance the LAION public dataset retained only about 10 percent of the total crawled data. In recent times data curation has gained prominence with several works developing strategies to retain high-quality subsets of raw scraped data. However these strategies are typically developed agnostic to the available compute for training. In this paper we demonstrate that making filtering decisions independent of training compute is often suboptimal: well-curated data rapidly loses its utility when repeated eventually decreasing below the utility of unseen but lower-quality data. While past research in neural scaling laws has considered web data to be homogenous real data is not. Our work bridges this important gap in the literature by developing scaling laws that characterize the differing utility of various data subsets and accounting for how this diminishes for a data point at its nth repetition. Our key message is that data curation can not be agnostic of the total compute a model will be trained for. Even without ever jointly training on multiple data buckets our scaling laws enable us to estimate model performance under this dynamic trade-off between quality and repetition. This allows us to curate the best possible pool for achieving top performance on Datacomp at various compute budgets carving out a pareto-frontier for data curation.

Computer vision techniques play a central role in the perception stack of autonomous vehicles. Such methods are employed to perceive the vehicle surroundings given sensor data. 3D LiDAR sensors are commonly used to collect sparse 3D point clouds from the scene. However compared to human perception such systems struggle to deduce the unseen parts of the scene given those sparse point clouds. In this matter the scene completion task aims at predicting the gaps in the LiDAR measurements to achieve a more complete scene representation. Given the promising results of recent diffusion models as generative models for images we propose extending them to achieve scene completion from a single 3D LiDAR scan. Previous works used diffusion models over range images extracted from LiDAR data directly applying image-based diffusion methods. Distinctly we propose to directly operate on the points reformulating the noising and denoising diffusion process such that it can efficiently work at scene scale. Together with our approach we propose a regularization loss to stabilize the noise predicted during the denoising process. Our experimental evaluation shows that our method can complete the scene given a single LiDAR scan as input producing a scene with more details compared to state-of-the-art scene completion methods. We believe that our proposed diffusion process formulation can support further research in diffusion models applied to scene-scale point cloud data.

Logit knowledge distillation attracts increasing attention due to its practicality in recent studies. However it often suffers inferior performance compared to the feature knowledge distillation. In this paper we argue that existing logit-based methods may be sub-optimal since they only leverage the global logit output that couples multiple semantic knowledge. This may transfer ambiguous knowledge to the student and mislead its learning. To this end we propose a simple but effective method i.e. Scale Decoupled Distillation (SDD) for logit knowledge distillation. SDD decouples the global logit output into multiple local logit outputs and establishes distillation pipelines for them. This helps the student to mine and inherit fine-grained and unambiguous logit knowledge. Moreover the decoupled knowledge can be further divided into consistent and complementary logit knowledge that transfers the semantic information and sample ambiguity respectively. By increasing the weight of complementary parts SDD can guide the student to focus more on ambiguous samples improving its discrimination ability. Extensive experiments on several benchmark datasets demonstrate the effectiveness of SDD for wide teacher-student pairs especially in the fine-grained classification task. Code is available at: \href https://github.com/shicaiwei123/SDD-CVPR2024 https://github.com/shicaiwei123/SDD-CVPR2024

Given an input set of 3D point pairs the goal of outlier-robust 3D registration is to compute some rotation and translation that align as many point pairs as possible. This is an important problem in computer vision for which many highly accurate approaches have been recently proposed. Despite their impressive performance these approaches lack scalability often overflowing the 16GB of memory of a standard laptop to handle roughly 30000 point pairs. In this paper we propose a 3D registration approach that can process more than ten million (10^7) point pairs with over 99% random outliers. Moreover our method is efficient entails low memory costs and maintains high accuracy at the same time. We call our method TEAR as it involves minimizing an outlier-robust loss that computes Truncated Entry-wise Absolute Residuals. To minimize this loss we decompose the original 6-dimensional problem into two subproblems of dimensions 3 and 2 respectively solved in succession to global optimality via a customized branch-and-bound method. While branch-and-bound is often slow and unscalable this does not apply to TEAR as we propose novel bounding functions that are tight and computationally efficient. Experiments on various datasets are conducted to validate the scalability and efficiency of our method.

In recent years semantic segmentation has become a pivotal tool in processing and interpreting satellite imagery. Yet a prevalent limitation of supervised learning techniques remains the need for extensive manual annotations by experts. In this work we explore the potential of generative image diffusion to address the scarcity of annotated data in earth observation tasks. The main idea is to learn the joint data manifold of images and labels leveraging recent advancements in denoising diffusion probabilistic models. To the best of our knowledge we are the first to generate both images and corresponding masks for satellite segmentation. We find that the obtained pairs not only display high quality in fine-scale features but also ensure a wide sampling diversity. Both aspects are crucial for earth observation data where semantic classes can vary severely in scale and occurrence frequency. We employ the novel data instances for downstream segmentation as a form of data augmentation. In our experiments we provide comparisons to prior works based on discriminative diffusion models or GANs. We demonstrate that integrating generated samples yields significant quantitative improvements for satellite semantic segmentation -- both compared to baselines and when training only on the original data.

DETR-like methods have significantly increased detection performance in an end-to-end manner. The mainstream two-stage frameworks of them perform dense self-attention and select a fraction of queries for sparse cross-attention which is proven effective for improving performance but also introduces a heavy computational burden and high dependence on stable query selection. This paper demonstrates that suboptimal two-stage selection strategies result in scale bias and redundancy due to the mismatch between selected queries and objects in two-stage initialization. To address these issues we propose hierarchical salience filtering refinement which performs transformer encoding only on filtered discriminative queries for a better trade-off between computational efficiency and precision. The filtering process overcomes scale bias through a novel scale-independent salience supervision. To compensate for the semantic misalignment among queries we introduce elaborate query refinement modules for stable two-stage initialization. Based on above improvements the proposed Salience DETR achieves significant improvements of +4.0% AP +0.2% AP +4.4% AP on three challenging task-specific detection datasets as well as 49.2% AP on COCO 2017 with less FLOPs. The code is available at https://github.com/xiuqhou/Salience-DETR.

Vision-language pre-training (VLP) aims to learn joint representations of vision and language modalities. The contrastive paradigm is currently dominant in this field. However we observe a notable misalignment phenomenon that is the affinity between samples has an obvious disparity across different modalities namely "Affinity Inconsistency Problem". Our intuition is that for a well-aligned model two images that look similar to each other should have the same level of similarity as their corresponding texts that describe them. In this paper we first investigate the reason of this inconsistency problem. We discover that the lack of consideration for sample-wise affinity consistency across modalities in existing training objectives is the central cause. To address this problem we propose a novel loss function named Sample-wise affinity Consistency (SaCo) loss which is designed to enhance such consistency by minimizing the distance between image embedding similarity and text embedding similarity for any two samples. Our SaCo loss can be easily incorporated into existing vision-language models as an additional loss due to its complementarity for most training objectives. In addition considering that pre-training from scratch is computationally expensive we also provide a more efficient way to continuously pre-train on a converged model by integrating our loss. Experimentally the model trained with our SaCo loss significantly outperforms the baseline on a variety of vision and language tasks.

Recently text-guided scalable vector graphics (SVGs) synthesis has shown promise in domains such as iconography and sketch. However existing text-to-SVG generation methods lack editability and struggle with visual quality and result diversity. To address these limitations we propose a novel text-guided vector graphics synthesis method called SVGDreamer. SVGDreamer incorporates a semantic-driven image vectorization (SIVE) process that enables the decomposition of synthesis into foreground objects and background thereby enhancing editability. Specifically the SIVE process introduces attention-based primitive control and an attention-mask loss function for effective control and manipulation of individual elements. Additionally we propose a Vectorized Particle-based Score Distillation (VPSD) approach to address issues of shape over-smoothing color over-saturation limited diversity and slow convergence of the existing text-to-SVG generation methods by modeling SVGs as distributions of control points and colors. Furthermore VPSD leverages a reward model to re-weight vector particles which improves aesthetic appeal and accelerates convergence. Extensive experiments are conducted to validate the effectiveness of SVGDreamer demonstrating its superiority over baseline methods in terms of editability visual quality and diversity. Project page: \href https://ximinng.github.io/SVGDreamer-project/ https://ximinng.github.io/SVGDreamer-project/

Tensor network (TN) representation is a powerful technique for computer vision and machine learning. TN structure search (TN-SS) aims to search for a customized structure to achieve a compact representation which is a challenging NP-hard problem. Recent "sampling-evaluation"-based methods require sampling an extensive collection of structures and evaluating them one by one resulting in prohibitively high computational costs. To address this issue we propose a novel TN paradigm named SVD-inspired TN decomposition (SVDinsTN) which allows us to efficiently solve the TN-SS problem from a regularized modeling perspective eliminating the repeated structure evaluations. To be specific by inserting a diagonal factor for each edge of the fully-connected TN SVDinsTN allows us to calculate TN cores and diagonal factors simultaneously with the factor sparsity revealing a compact TN structure. In theory we prove a convergence guarantee for the proposed method. Experimental results demonstrate that the proposed method achieves approximately 100 1000 times acceleration compared to the state-of-the-art TN-SS methods while maintaining a comparable level of representation ability.

Efficiently representing and reconstructing the 3D geometry of biological trees remains a challenging problem in computer vision and graphics. We propose a novel approach for generating realistic tree models from single-view photographs. We cast the 3D information inference problem to a semantic voxel diffusion process which converts an input image of a tree to a novel Semantic Voxel Structure (SVS) in 3D space. The SVS encodes the geometric appearance and semantic structural information (e.g. classifying trunks branches and leaves) which retains the intricate internal tree features. Tailored to the SVS we present SVDTree a new hybrid tree modeling approach by combining structure-oriented branch reconstruction and self-organization-based foliage reconstruction. We validate SVDTree by using images from both synthetic and real trees. The comparison results show that our approach can better preserve tree details and achieve more realistic and accurate reconstruction results than previous methods.

In this paper we revisit techniques for uncertainty estimation within deep neural networks and consolidate a suite of techniques to enhance their reliability. Our investigation reveals that an integrated application of diverse techniques--spanning model regularization classifier and optimization--substantially improves the accuracy of uncertainty predictions in image classification tasks. The synergistic effect of these techniques culminates in our novel SURE approach. We rigorously evaluate SURE against the benchmark of failure prediction a critical testbed for uncertainty estimation efficacy. Our results showcase a consistently better performance than models that individually deploy each technique across various datasets and model architectures. When applied to real-world challenges such as data corruption label noise and long-tailed class distribution SURE exhibits remarkable robustness delivering results that are superior or on par with current state-of-the-art specialized methods. Particularly on Animal-10N and Food-101N for learning with noisy labels SURE achieves state-of-the-art performance without any task-specific adjustments. This work not only sets a new benchmark for robust uncertainty estimation but also paves the way for its application in diverse real-world scenarios where reliability is paramount. Our code is available at https://yutingli0606.github.io/SURE/.

Learning generalizable visual representations from Internet data has yielded promising results for robotics. Yet prevailing approaches focus on pre-training 2D representations being sub-optimal to deal with occlusions and accurately localize objects in complex 3D scenes. Meanwhile 3D representation learning has been limited to single-object understanding. To address these limitations we introduce a novel 3D pre-training framework for robotics named SUGAR that captures semantic geometric and affordance properties of objects through 3D point clouds. We underscore the importance of cluttered scenes in 3D representation learning and automatically construct a multi-object dataset benefiting from cost-free supervision in simulation. SUGAR employs a versatile transformer-based model to jointly address five pre-training tasks namely cross-modal knowledge distillation for semantic learning masked point modeling to understand geometry structures grasping pose synthesis for object affordance 3D instance segmentation and referring expression grounding to analyze cluttered scenes. We evaluate our learned representation on three robotic-related tasks namely zero-shot 3D object recognition referring expression grounding and language-driven robotic manipulation. Experimental results show that SUGAR's 3D representation outperforms state-of-the-art 2D and 3D representations.

Recent advancements in subject-driven image generation have led to zero-shot generation yet precise selection and focus on crucial subject representations remain challenging. Addressing this we introduce the SSR-Encoder a novel architecture designed for selectively capturing any subject from single or multiple reference images. It responds to various query modalities including text and masks without necessitating test-time fine-tuning. The SSR-Encoder combines a Token-to-Patch Aligner that aligns query inputs with image patches and a Detail-Preserving Subject Encoder for extracting and preserving fine features of the subjects thereby generating subject embeddings. These embeddings used in conjunction with original text embeddings condition the generation process. Characterized by its model generalizability and efficiency the SSR-Encoder adapts to a range of custom models and control modules. Enhanced by the Embedding Consistency Regularization Loss for improved training our extensive experiments demonstrate its effectiveness in versatile and high-quality image generation indicating its broad applicability.

In recent years large-scale video-language pre-training (VidLP) has received considerable attention for its effectiveness in relevant tasks. In this paper we propose a novel action-centric VidLP framework that employs video tube features for temporal modeling and language features based on semantic role labeling (SRL). Our video encoder generates multiple tube features along object trajectories identifying action-related regions within videos to overcome the limitations of existing temporal attention mechanisms. Additionally our text encoder incorporates high-level action-related language knowledge previously underutilized in current VidLP models. The SRL captures action-verbs and related semantics among objects in sentences and enhances the ability to perform instance-level text matching thus enriching the cross-modal (CM) alignment process. We also introduce two novel pre-training objectives and a self-supervision strategy to produce a more faithful CM representation. Experimental results demonstrate that our method outperforms existing VidLP frameworks in various downstream tasks and datasets establishing our model a baseline in the modern VidLP framework.

Despite the success of recent upsampling approaches generating high-resolution point sets with uniform distribution and meticulous structures is still challenging. Unlike existing methods that only take spatial information of the raw data into account we regard point cloud upsampling as generating dense point clouds from deformable topology. Motivated by this we present SPU-PMD a self-supervised topological mesh deformation network for 3D densification. As a cascaded framework our architecture is formulated by a series of coarse mesh interpolator and mesh deformers. At each stage the mesh interpolator first produces the initial dense point clouds via mesh interpolation which allows the model to perceive the primitive topology better. Meanwhile the deformer infers the morphing by estimating the movements of mesh nodes and reconstructs the descriptive topology structure. By associating mesh deformation with feature expansion this module progressively refines point clouds' surface uniformity and structural details. To demonstrate the effectiveness of the proposed method extensive quantitative and qualitative experiments are conducted on synthetic and real-scanned 3D data. Also we compare it with state-of-the-art techniques to further illustrate the superiority of our network. The project page is: https://github.com/lyz21/SPU-PMD

Unsupervised object-centric learning aims to decompose scenes into interpretable object entities termed slots. Slot-based auto-encoders stand out as a prominent method for this task. Within them crucial aspects include guiding the encoder to generate object-specific slots and ensuring the decoder utilizes them during reconstruction. This work introduces two novel techniques (i) an attention-based self-training approach which distills superior slot-based attention masks from the decoder to the encoder enhancing object segmentation and (ii) an innovative patch-order permutation strategy for autoregressive transformers that strengthens the role of slot vectors in reconstruction. The effectiveness of these strategies is showcased experimentally. The combined approach significantly surpasses prior slot-based autoencoder methods in unsupervised object segmentation especially with complex real-world images. We provide the implementation code at https://github.com/gkakogeorgiou/spot .

Reinforcement learning (RL) with dense rewards and imitation learning (IL) with human-generated trajectories are the most widely used approaches for training modern embodied agents. RL requires extensive reward shaping and auxiliary losses and is often too slow and ineffective for long-horizon tasks. While IL with human supervision is effective collecting human trajectories at scale is extremely expensive. In this work we show that imitating shortest-path planners in simulation produces agents that given a language instruction can proficiently navigate explore and manipulate objects in both simulation and in the real world using only RGB sensors (no depth map or GPS coordinates). This surprising result is enabled by our end-to-end transformer-based SPOC architecture powerful visual encoders paired with extensive image augmentation and the dramatic scale and diversity of our training data: millions of frames of shortest-path-expert trajectories collected inside approximately 200000 procedurally generated houses containing 40000 unique 3D assets. Our models data training code and newly proposed 10-task benchmarking suite CHORES are available at https://spoc-robot.github.io/.