CVPR 2024 Papers

Vision-language pre-trained models have achieved impressive performance on various downstream tasks. However their large model sizes hinder their utilization on platforms with limited computational resources. We find that directly using smaller pre-trained models and applying magnitude-based pruning on CLIP models leads to inflexibility and inferior performance. Recent efforts for VLP compression either adopt uni-modal compression metrics resulting in limited performance or involve costly mask-search processes with learnable masks. In this paper we first propose the Module-wise Pruning Error (MoPE) metric accurately assessing CLIP module importance by performance decline on cross-modal tasks. Using the MoPE metric we introduce a unified pruning framework applicable to both pre-training and task-specific fine-tuning compression stages. For pre-training MoPE-CLIP effectively leverages knowledge from the teacher model significantly reducing pre-training costs while maintaining strong zero-shot capabilities. For fine-tuning consecutive pruning from width to depth yields highly competitive task-specific models. Extensive experiments in two stages demonstrate the effectiveness of the MoPE metric and MoPE-CLIP outperforms previous state-of-the-art VLP compression methods.

We introduce MoMask a novel masked modeling framework for text-driven 3D human motion generation. In MoMask a hierarchical quantization scheme is employed to represent human motion as multi-layer discrete motion tokens with high-fidelity details. Starting at the base layer with a sequence of motion tokens obtained by vector quantization the residual tokens of increasing orders are derived and stored at the subsequent layers of the hierarchy. This is consequently followed by two distinct bidirectional transformers. For the base-layer motion tokens a Masked Transformer is designated to predict randomly masked motion tokens conditioned on text input at training stage. During generation (i.e. inference) stage starting from an empty sequence our Masked Transformer iteratively fills up the missing tokens; Subsequently a Residual Transformer learns to progressively predict the next-layer tokens based on the results from current layer. Extensive experiments demonstrate that MoMask outperforms the state-of-art methods on the text-to-motion generation task with an FID of 0.045 (vs e.g. 0.141 of T2M-GPT) on the HumanML3D dataset and 0.228 (vs 0.514) on KIT-ML respectively. MoMask can also be seamlessly applied in related tasks without further model fine-tuning such as text-guided temporal inpainting.

In the academic field the research on human motion prediction tasks mainly focuses on exploiting the observed information to forecast human movements accurately in the near future horizon. However a significant gap appears when it comes to the application field as current models are all trained offline with fixed parameters that are inherently suboptimal to handle the complex yet ever-changing nature of human behaviors. To bridge this gap in this paper we introduce the task of online meta adaptation for human motion prediction based on the insight that finding "smart weights" capable of swift adjustments to suit different motion contexts along the time is a key to improving predictive accuracy. We propose MoML which ingeniously borrows the bilevel optimization spirit of model-agnostic meta-learning to transform previous predictive mistakes into strong inductive biases to guide online adaptation. This is achieved by our MoAdapter blocks that can learn error information by facilitating efficient adaptation via a few gradient steps which fine-tunes our meta-learned "smart" initialization produced by the generic predictor. Considering real-time requirements in practice we further propose Fast-MoML a more efficient variant of MoML that features a closed-form solution instead of conventional gradient update. Experimental results show that our approach can effectively bring many existing offline motion prediction models online and improves their predictive accuracy.

The success of contrastive language-image pretraining (CLIP) relies on the supervision from the pairing between images and captions which tends to be noisy in web-crawled data. We present Mixture of Data Experts (MoDE) and learn a system of CLIP data experts via clustering. Each data expert is trained on one data cluster being less sensitive to false negative noises in other clusters. At inference time we ensemble their outputs by applying weights determined through the correlation between task metadata and cluster conditions. To estimate the correlation precisely the samples in one cluster should be semantically similar but the number of data experts should still be reasonable for training and inference. As such we consider the ontology in human language and propose to use fine-grained cluster centers to represent each data expert at a coarse-grained level. Experimental studies show that four CLIP data experts on ViT-B/16 outperform the ViT-L/14 by OpenAI CLIP and OpenCLIP on zero-shot image classification but with less (<35%) training cost. Meanwhile MoDE can train all data expert asynchronously and can flexibly include new data experts. The code is available here.

Learning-based stereo matching techniques have made significant progress. However existing methods inevitably lose geometrical structure information during the feature channel generation process resulting in edge detail mismatches. In this paper the Motif Channel Attention Stereo Matching Network (MoCha-Stereo) is designed to address this problem. We provide the Motif Channel Correlation Volume (MCCV) to determine more accurate edge matching costs. MCCV is achieved by projecting motif channels which capture common geometric structures in feature channels onto feature maps and cost volumes. In addition edge variations in the reconstruction error map also affect details matching we propose the Reconstruction Error Motif Penalty (REMP) module to further refine the full-resolution disparity estimation. REMP integrates the frequency information of typical channel features from the reconstruction error. MoCha-Stereo ranks 1st on the KITTI-2015 and KITTI-2012 Reflective leaderboards. Our structure also shows excellent performance in Multi-View Stereo. Code is avaliable at https://github.com/ZYangChen/MoCha-Stereo.

While federated learning (FL) systems often utilize quantization to battle communication and computational bottlenecks they have heretofore been limited to deploying fixed-precision quantization schemes. Meanwhile the concept of mixed-precision quantization (MPQ) where different layers of a deep learning model are assigned varying bit-width remains unexplored in the FL settings. We present a novel FL algorithm FedMPQ which introduces mixed-precision quantization to resource-heterogeneous FL systems. Specifically local models quantized so as to satisfy bit-width constraint are trained by optimizing an objective function that includes a regularization term which promotes reduction of precision in some of the layers without significant performance degradation. The server collects local model updates de-quantizes them into full-precision models and then aggregates them into a global model. To initialize the next round of local training the server relies on the information learned in the previous training round to customize bit-width assignments of the models delivered to different clients. In extensive benchmarking experiments on several model architectures and different datasets in both iid and non-iid settings FedMPQ outperformed the baseline FL schemes that utilize fixed-precision quantization while incurring only a minor computational overhead on the participating devices.

Large Vision-Language Models (LVLMs) have advanced considerably intertwining visual recognition and language understanding to generate content that is not only coherent but also contextually attuned. Despite their success LVLMs still suffer from the issue of object hallucinations where models generate plausible yet incorrect outputs that include objects that do not exist in the images. To mitigate this issue we introduce Visual Contrastive Decoding (VCD) a simple and training-free method that contrasts output distributions derived from original and distorted visual inputs. The proposed VCD effectively reduces the over-reliance on statistical bias and unimodal priors two essential causes of object hallucinations. This adjustment ensures the generated content is closely grounded to visual inputs resulting in contextually accurate outputs. Our experiments show that VCD without either additional training or the usage of external tools significantly mitigates the object hallucination issue across different LVLM families. Beyond mitigating object hallucinations VCD also excels in general LVLM benchmarks highlighting its wide-ranging applicability.

In the realm of point cloud scene understanding particularly in indoor scenes objects are arranged following human habits resulting in objects of certain semantics being closely positioned and displaying notable inter-object correlations. This can create a tendency for neural networks to exploit these strong dependencies bypassing the individual object patterns. To address this challenge we introduce a novel self-supervised learning (SSL) strategy. Our approach leverages both object patterns and contextual cues to produce robust features. It begins with the formulation of an object-exchanging strategy where pairs of objects with comparable sizes are exchanged across different scenes effectively disentangling the strong contextual dependencies. Subsequently we introduce a context-aware feature learning strategy which encodes object patterns without relying on their specific context by aggregating object features across various scenes. Our extensive experiments demonstrate the superiority of our method over existing SSL techniques further showing its better robustness to environmental changes. Moreover we showcase the applicability of our approach by transferring pre-trained models to diverse point cloud datasets.

Noisy correspondence that refers to mismatches in cross-modal data pairs is prevalent on human-annotated or web-crawled datasets. Prior approaches to leverage such data mainly consider the application of uni-modal noisy label learning without amending the impact on both cross-modal and intra-modal geometrical structures in multimodal learning. Actually we find that both structures are effective to discriminate noisy correspondence through structural differences when being well-established. Inspired by this observation we introduce a Geometrical Structure Consistency (GSC) method to infer the true correspondence. Specifically GSC ensures the preservation of geometrical structures within and between modalities allowing for the accurate discrimination of noisy samples based on structural differences. Utilizing these inferred true correspondence labels GSC refines the learning of geometrical structures by filtering out the noisy samples. Experiments across four cross-modal datasets confirm that GSC effectively identifies noisy samples and significantly outperforms the current leading methods. Source code is available at https://github.com/MediaBrain-SJTU/GSC.

Neural Radiance Fields (NeRFs) have shown great potential in novel view synthesis. However they struggle to render sharp images when the data used for training is affected by motion blur. On the other hand event cameras excel in dynamic scenes as they measure brightness changes with microsecond resolution and are thus only marginally affected by blur. Recent methods attempt to enhance NeRF reconstructions under camera motion by fusing frames and events. However they face challenges in recovering accurate color content or constrain the NeRF to a set of predefined camera poses harming reconstruction quality in challenging conditions. This paper proposes a novel formulation addressing these issues by leveraging both model- and learning-based modules. We explicitly model the blur formation process exploiting the event double integral as an additional model-based prior. Additionally we model the event-pixel response using an end-to-end learnable response function allowing our method to adapt to non-idealities in the real event-camera sensor. We show on synthetic and real data that the proposed approach outperforms existing deblur NeRFs that use only frames as well as those that combine frames and events by +6.13dB and +2.48dB respectively.

This paper aims to address a common challenge in deep learning-based image transformation methods such as image enhancement and super-resolution which heavily rely on precisely aligned paired datasets with pixel-level alignments. However creating precisely aligned paired images presents significant challenges and hinders the advancement of methods trained on such data. To overcome this challenge this paper introduces a novel and simple Frequency Distribution Loss (FDL) for computing distribution distance within the frequency domain. Specifically we transform image features into the frequency domain using Discrete Fourier Transformation (DFT). Subsequently frequency components (amplitude and phase) are processed separately to form the FDL loss function. Our method is empirically proven effective as a training constraint due to the thoughtful utilization of global information in the frequency domain. Extensive experimental evaluations focusing on image enhancement and super-resolution tasks demonstrate that FDL outperforms existing misalignment-robust loss functions. Furthermore we explore the potential of our FDL for image style transfer that relies solely on completely misaligned data. Our code is available at: https://github.com/eezkni/FDL

One of the main challenges of multimodal learning is the need to combine heterogeneous modalities (e.g. video audio text). For example video and audio are obtained at much higher rates than text and are roughly aligned in time. They are often not synchronized with text which comes as a global context e.g. a title or a description. Furthermore video and audio inputs are of much larger volumes and grow as the video length increases which naturally requires more compute dedicated to these modalities and makes modeling of long-range dependencies harder. We here decouple the multimodal modeling dividing it into separate autoregressive models processing the inputs according to the characteristics of the modalities. We propose a multimodal model consisting of an autoregressive component for the time-synchronized modalities (audio and video) and an autoregressive component for the context modalities which are not necessarily aligned in time but are still sequential. To address the long-sequences of the video-audio inputs we further partition the video and audio sequences in consecutive snippets and autoregressively process their representations. To that end we propose a Combiner mechanism which models the audio-video information jointly producing compact but expressive representations. This allows us to scale to 512 input video frames without increase in model parameters. Our approach achieves the state-of-the-art on multiple well established multimodal benchmarks. It effectively addresses the high computational demand of media inputs by learning compact representations controlling the sequence length of the audio-video feature representations and modeling their dependencies in time.

Nowadays leveraging 2D images and pre-trained models to guide 3D point cloud feature representation has shown a remarkable potential to boost the performance of 3D fundamental models. While some works rely on additional data such as 2D real-world images and their corresponding camera poses recent studies target at using point cloud exclusively by designing 3D-to-2D projection. However in the indoor scene scenario existing 3D-to-2D projection strategies suffer from severe occlusions and incoherence which fail to contain sufficient information for fine-grained point cloud segmentation task. In this paper we argue that the crux of the matter resides in the basic premise of existing projection strategies that the medium is homogeneous thereby projection rays propagate along straight lines and behind objects are occluded by front ones. Inspired by the phenomenon of mirage where the occluded objects are exposed by distorted light rays due to heterogeneous medium refraction rate we propose MirageRoom by designing parametric mirage projection with heterogeneous medium to obtain series of projected images with various distorted degrees. We further develop a masked reprojection module across 2D and 3D latent space to bridge the gap between pre-trained 2D backbone and 3D point-wise features. Both quantitative and qualitative experimental results on S3DIS and ScanNet V2 demonstrate the effectiveness of our method.

Recently 3D Gaussian Splatting has demonstrated impressive novel view synthesis results reaching high fidelity and efficiency. However strong artifacts can be observed when changing the sampling rate e.g. by changing focal length or camera distance. We find that the source for this phenomenon can be attributed to the lack of 3D frequency constraints and the usage of a 2D dilation filter. To address this problem we introduce a 3D smoothing filter to constrains the size of the 3D Gaussian primitives based on the maximal sampling frequency induced by the input views. It eliminates high-frequency artifacts when zooming in. Moreover replacing 2D dilation with a 2D Mip filter which simulates a 2D box filter effectively mitigates aliasing and dilation issues. Our evaluation including scenarios such a training on single-scale images and testing on multiple scales validates the effectiveness of our approach.

This paper focuses on self-supervised monocular depth estimation in dynamic scenes trained on monocular videos. Existing methods jointly estimate pixel-wise depth and motion relying mainly on an image reconstruction loss. Dynamic regions remain a critical challenge for these methods due to the inherent ambiguity in depth and motion estimation resulting in inaccurate depth estimation. This paper proposes a self-supervised training framework exploiting pseudo depth labels for dynamic regions from training data. The key contribution of our framework is to decouple depth estimation for static and dynamic regions of images in the training data. We start with an unsupervised depth estimation approach which provides reliable depth estimates for static regions and motion cues for dynamic regions and allows us to extract moving object information at the instance level. In the next stage we use an object network to estimate the depth of those moving objects assuming rigid motions. Then we propose a new scale alignment module to address the scale ambiguity between estimated depths for static and dynamic regions. We can then use the depth labels generated to train an end-to-end depth estimation network and improve its performance. Extensive experiments on the Cityscapes and KITTI datasets show that our self-training strategy consistently outperforms existing self-/unsupervised depth estimation methods.

We introduce a new family of minimal problems for reconstruction from multiple views. Our primary focus is a novel approach to autocalibration a long-standing problem in computer vision. Traditional approaches to this problem such as those based on Kruppa's equations or the modulus constraint rely explicitly on the knowledge of multiple fundamental matrices or a projective reconstruction. In contrast we consider a novel formulation involving constraints on image points the unknown depths of 3D points and a partially specified calibration matrix K. For 2 and 3 views we present a comprehensive taxonomy of minimal autocalibration problems obtained by relaxing some of these constraints. These problems are organized into classes according to the number of views and any assumed prior knowledge of K. Within each class we determine problems with the fewest---or a relatively small number of---solutions. From this zoo of problems we devise three practical solvers. Experiments with synthetic and real data and interfacing our solvers with COLMAP demonstrate that we achieve superior accuracy compared to state-of-the-art calibration methods. The code is available at https://github.com/andreadalcin/MinimalPerspectiveAutocalibration.

Brain decoding a pivotal field in neuroscience aims to reconstruct stimuli from acquired brain signals primarily utilizing functional magnetic resonance imaging (fMRI). Currently brain decoding is confined to a per-subject-per-model paradigm limiting its applicability to the same individual for whom the decoding model is trained. This constraint stems from three key challenges: 1) the inherent variability in input dimensions across subjects due to differences in brain size; 2) the unique intrinsic neural patterns influencing how different individuals perceive and process sensory information; 3) limited data availability for new subjects in real-world scenarios hampers the performance of decoding models. In this paper we present a novel approach MindBridge that achieves cross-subject brain decoding by employing only one model. Our proposed framework establishes a generic paradigm capable of addressing these challenges by introducing biological-inspired aggregation function and novel cyclic fMRI reconstruction mechanism for subject-invariant representation learning. Notably by cycle reconstruction of fMRI MindBridge can enable novel fMRI synthesis which also can serve as pseudo data augmentation. Within the framework we also devise a novel reset-tuning method for adapting a pretrained model to a new subject. Experimental results demonstrate MindBridge's ability to reconstruct images for multiple subjects which is competitive with dedicated subject-specific models. Furthermore with limited data for a new subject we achieve a high level of decoding accuracy surpassing that of subject-specific models. This advancement in cross-subject brain decoding suggests promising directions for wider applications in neuroscience and indicates potential for more efficient utilization of limited fMRI data in real-world scenarios. Project page: https://littlepure2333.github.io/MindBridge

Monocular Depth Estimation (MDE) is a fundamental problem in computer vision with numerous applications. Recently LIDAR-supervised methods have achieved remarkable per-pixel depth accuracy in outdoor scenes. However significant errors are typically found in the proximity of depth discontinuities i.e. depth edges which often hinder the performance of depth-dependent applications that are sensitive to such inaccuracies e.g. novel view synthesis and augmented reality. Since direct supervision for the location of depth edges is typically unavailable in sparse LIDAR-based scenes encouraging the MDE model to produce correct depth edges is not straightforward. To the best of our knowledge this paper is the first attempt to address the depth edges issue for LIDAR-supervised scenes. In this work we propose to learn to detect the location of depth edges from densely-supervised synthetic data and use it to generate supervision for the depth edges in the MDE training. To quantitatively evaluate our approach and due to the lack of depth edges GT in LIDAR-based scenes we manually annotated subsets of the KITTI and the DDAD datasets with depth edges ground truth. We demonstrate significant gains in the accuracy of the depth edges with comparable per-pixel depth accuracy on several challenging datasets. Code and datasets are available at https://github.com/liortalker/MindTheEdge.

Crack segmentation datasets make great efforts to obtain the ground truth crack or non-crack labels as clearly as possible. However it can be observed that ambiguities are still inevitable when considering the marginal non-crack region due to low contrast and heterogeneous texture. To solve this problem we propose a novel clustering-inspired representation learning framework which contains a two-phase strategy for automatic crack segmentation. In the first phase a pre-process is proposed to localize the marginal non-crack region. Then we propose an ambiguity-aware segmentation loss (Aseg Loss) that enables crack segmentation models to capture ambiguities in the above regions via learning segmentation variance which allows us to further localize ambiguous regions. In the second phase to learn the discriminative features of the above regions we propose a clustering-inspired loss (CI Loss) that alters the supervision learning of these regions into an unsupervised clustering manner. We demonstrate that the proposed method could surpass the existing crack segmentation models on various datasets and our constructed CrackSeg5k dataset.