CVPR 2024 Papers

The rapid evolution of automatic facial indexing tech- nologies increases the risk of compromising personal and sensitive information. To address the issue we propose cre- ating cartoon avatars or 'toon avatars' designed to effec- tively obscure identity features. The primary objective is to deceive current AI systems preventing them from accu- rately identifying individuals while making minimal modi- fications to their facial features. Moreover we aim to en- sure that a human observer can still recognize the person depicted in these altered avatar images. To achieve this we introduce 'ToonerGAN' a novel approach that utilizes Generative Adversarial Networks (GANs) to craft person- alized cartoon avatars. The ToonerGAN framework con- sists of a style module and a de-identification module that work together to produce high-resolution realistic cartoon images. For the efficient training of our network we have developed an extensive dataset named 'ToonSet' compris- ing approximately 23000 facial images and their cartoon renditions. Through comprehensive experiments and bench- marking against existing datasets including CelebA-HQ our method demonstrates superior performance in obfus- cating identity while preserving the utility of data. Addi- tionally a user-centric study to explore the effectiveness of ToonerGAN has yielded some compelling observations.

We address the problem of regressing 3D human pose and shape from a single image with a focus on 3D accuracy. The current best methods leverage large datasets of 3D pseudo-ground-truth (p-GT) and 2D keypoints leading to robust performance. With such methods however we observe a paradoxical decline in 3D pose accuracy with increasing 2D accuracy. This is caused by biases in the p-GT and the use of an approximate camera projection model. We quantify the error induced by current camera models and show that fitting 2D keypoints and p-GT accurately causes incorrect 3D poses. Our analysis defines the invalid distances within which minimizing 2D and p-GT losses is detrimental. We use this to formulate a new loss "Threshold-Adaptive Loss Scaling" (TALS) that penalizes gross 2D and p-GT errors but not smaller ones. With such a loss there are many 3D poses that could equally explain the 2D evidence. To reduce this ambiguity we need a prior over valid human poses but such priors can introduce unwanted bias. To address this we exploit a tokenized representation of human pose and reformulate the problem as token prediction. This restricts the estimated poses to the space of valid poses effectively improving robustness to occlusion. Extensive experiments on the EMDB and 3DPW datasets show that our reformulated loss and tokenization allows us to train on in-the-wild data while improving 3D accuracy over the state-of-the-art. Our models and code are available for research at https://tokenhmr.is.tue.mpg.de.

We present TokenCompose a Latent Diffusion Model for text-to-image generation that achieves enhanced consistency between user-specified text prompts and model-generated images. Despite its tremendous success the standard denoising process in the Latent Diffusion Model takes text prompts as conditions only absent explicit constraint for the consistency between the text prompts and the image contents leading to unsatisfactory results for composing multiple object categories. Our proposed TokenCompose aims to improve multi-category instance composition by introducing the token-wise consistency terms between the image content and object segmentation maps in the finetuning stage. TokenCompose can be applied directly to the existing training pipeline of text-conditioned diffusion models without extra human labeling information. By finetuning Stable Diffusion with our approach the model exhibits significant improvements in multi-category instance composition and enhanced photorealism for its generated images.

While Transformers have rapidly gained popularity in various computer vision applications post-hoc explanations of their internal mechanisms remain largely unexplored. Vision Transformers extract visual information by representing image regions as transformed tokens and integrating them via attention weights. However existing post-hoc explanation methods merely consider these attention weights neglecting crucial information from the transformed tokens which fails to accurately illustrate the rationales behind the models' predictions. To incorporate the influence of token transformation into interpretation we propose TokenTM a novel post-hoc explanation method that utilizes our introduced measurement of token transformation effects. Specifically we quantify token transformation effects by measuring changes in token lengths and correlations in their directions pre- and post-transformation. Moreover we develop initialization and aggregation rules to integrate both attention weights and token transformation effects across all layers capturing holistic token contributions throughout the model. Experimental results on segmentation and perturbation tests demonstrate the superiority of our proposed TokenTM compared to state-of-the-art Vision Transformer explanation methods.

Annotating lots of 3D medical images for training segmentation models is time-consuming. The goal of weakly supervised semantic segmentation is to train segmentation models without using any ground truth segmentation masks. Our work addresses the case where only image-level categorical labels indicating the presence or absence of a particular region of interest (such as tumours or lesions) are available. Most existing methods rely on class activation mapping (CAM). We propose a novel approach ToNNO which is based on the Tomographic reconstruction of a Neural Network's Output. Our technique extracts stacks of slices with different angles from the input 3D volume feeds these slices to a 2D encoder and applies the inverse Radon transform in order to reconstruct a 3D heatmap of the encoder's predictions. This generic method allows to perform dense prediction tasks on 3D volumes using any 2D image encoder. We apply it to weakly supervised medical image segmentation by training the 2D encoder to output high values for slices containing the regions of interest. We test it on four large scale medical image datasets and outperform 2D CAM methods. We then extend ToNNO by combining tomographic reconstruction with CAM methods proposing Averaged CAM and Tomographic CAM which obtain even better results.

This work proposes TimeChat a time-sensitive multimodal large language model specifically designed for long video understanding. Our model incorporates two key architectural contributions: (1) a timestamp-aware frame encoder that binds visual content with the timestamp of each frame and (2) a sliding video Q-Former that produces a video token sequence of varying lengths to accommodate videos of various durations. Additionally we construct an instruction-tuning dataset encompassing 6 tasks and a total of 125K instances to further enhance TimeChat's instruction-following performance. Experiment results across various video understanding tasks such as dense captioning temporal grounding and highlight detection demonstrate TimeChat's strong zero-shot temporal localization and reasoning capabilities. For example it achieves +9.2 F1 score and +2.8 CIDEr on YouCook2 +5.8 HIT@1 on QVHighlights and +27.5 R@1 (IoU=0.5) on Charades-STA compared to state-of-the-art video large language models holding the potential to serve as a versatile video assistant for long-form video comprehension tasks and satisfy realistic user requirements.

We propose a computational imaging method for time-efficient light-field acquisition that combines a coded aperture with an event-based camera. Different from the conventional coded-aperture imaging method our method applies a sequence of coding patterns during a single exposure for an image frame. The parallax information which is related to the differences in coding patterns is recorded as events. The image frame and events all of which are measured in a single exposure are jointly used to computationally reconstruct a light field. We also designed an algorithm pipeline for our method that is end-to-end trainable on the basis of deep optics and compatible with real camera hardware. We experimentally showed that our method can achieve more accurate reconstruction than several other imaging methods with a single exposure. We also developed a hardware prototype with the potential to complete the measurement on the camera within 22 msec and demonstrated that light fields from real 3-D scenes can be obtained with convincing visual quality. Our software and supplementary video are available from our project website.

As foundation models become more popular there is a growing need to efficiently finetune them for downstream tasks. Although numerous adaptation methods have been proposed they are designed to be efficient only in terms of how many parameters are trained. They however typically still require backpropagating gradients throughout the model meaning that their training-time and -memory cost does not reduce as significantly. We propose an adaptation method which does not backpropagate gradients through the backbone. We achieve this by designing a lightweight network in parallel that operates on features from the frozen pretrained backbone. As a result our method is efficient not only in terms of parameters but also in training-time and memory usage. Our approach achieves state-of-the-art accuracy-parameter trade-offs on the popular VTAB benchmark and we further show how we outperform prior works with respect to training-time and -memory usage too. We further demonstrate the training efficiency and scalability of our method by adapting a vision transformer backbone of 4 billion parameters for the computationally demanding task of video classification without any intricate model parallelism. Here we outperform a prior adaptor-based method which could only scale to a 1 billion parameter backbone or fully-finetuning a smaller backbone with the same GPU and less training time.

Despite many attempts to leverage pre-trained text-to-image models (T2I) like Stable Diffusion (SD) for controllable image editing producing good predictable results remains a challenge. Previous approaches have focused on either fine-tuning pre-trained T2I models on specific datasets to generate certain kinds of images (e.g. with a specific object or person) or on optimizing the weights text prompts and/or learning features for each input image in an attempt to coax the image generator to produce the desired result. However these approaches all have shortcomings and fail to produce good results in a predictable and controllable manner. To address this problem we present TiNO-Edit an SD-based method that focuses on optimizing the noise patterns and diffusion timesteps during editing something previously unexplored in the literature. With this simple change we are able to generate results that both better align with the original images and reflect the desired result. Furthermore we propose a set of new loss functions that operate in the latent domain of SD greatly speeding up the optimization when compared to prior losses which operate in the pixel domain. Our method can be easily applied to variations of SD including Textual Inversion and DreamBooth that encode new concepts and incorporate them into the edited results. We present a host of image-editing capabilities enabled by our approach. Our code is publicly available at https://github.com/SherryXTChen/TiNO-Edit.

Self-supervised image backbones can be used to address complex 2D tasks (e.g. semantic segmentation object discovery) very efficiently and with little or no downstream supervision. Ideally 3D backbones for lidar should be able to inherit these properties after distillation of these powerful 2D features. The most recent methods for image-to-lidar distillation on autonomous driving data show promising results obtained thanks to distillation methods that keep improving. Yet we still notice a large performance gap when measuring by linear probing the quality of distilled vs fully supervised features. In this work instead of focusing only on the distillation method we study the effect of three pillars for distillation: the 3D backbone the pretrained 2D backbone and the pretraining 2D+3D dataset. In particular thanks to our scalable distillation method named ScaLR we show that scaling the 2D and 3D backbones and pretraining on diverse datasets leads to a substantial improvement of the feature quality. This allows us to significantly reduce the gap between the quality of distilled and fully-supervised 3D features and to improve the robustness of the pretrained backbones to domain gaps and perturbations.

Federated learning facilitates the collaborative learning of a global model across multiple distributed medical institutions without centralizing data. Nevertheless the expensive cost of annotation on local clients remains an obstacle to effectively utilizing local data. To mitigate this issue federated active learning methods suggest leveraging local and global model predictions to select a relatively small amount of informative local data for annotation. However existing methods mainly focus on all local data sampled from the same domain making them unreliable in realistic medical scenarios with domain shifts among different clients. In this paper we make the first attempt to assess the informativeness of local data derived from diverse domains and propose a novel methodology termed Federated Evidential Active Learning (FEAL) to calibrate the data evaluation under domain shift. Specifically we introduce a Dirichlet prior distribution in both local and global models to treat the prediction as a distribution over the probability simplex and capture both aleatoric and epistemic uncertainties by using the Dirichlet-based evidential model. Then we employ the epistemic uncertainty to calibrate the aleatoric uncertainty. Afterward we design a diversity relaxation strategy to reduce data redundancy and maintain data diversity. Extensive experiments and analysis on five real multi-center medical image datasets demonstrate the superiority of FEAL over the state-of-the-art active learning methods in federated scenarios with domain shifts. The code will be available at https://github.com/JiayiChen815/FEAL.

Considerable efforts have been devoted to Oriented Object Detection (OOD). However one lasting issue regarding the discontinuity in Oriented Bounding Box (OBB) representation remains unresolved which is an inherent bottleneck for extant OOD methods. This paper endeavors to completely solve this issue in a theoretically guaranteed manner and puts an end to the ad-hoc efforts in this direction. Prior studies typically can only address one of the two cases of discontinuity: rotation and aspect ratio and often inadvertently introduce decoding discontinuity e.g. Decoding Incompleteness (DI) and Decoding Ambiguity (DA) as discussed in literature. Specifically we propose a novel representation method called Continuous OBB (COBB) which can be readily integrated into existing detectors e.g. Faster-RCNN as a plugin. It can theoretically ensure continuity in bounding box regression which to our best knowledge has not been achieved in literature for rectangle-based object representation. For fairness and transparency of experiments we have developed a modularized benchmark based on the open-source deep learning framework Jittor's detection toolbox JDet for OOD evaluation. On the popular DOTA dataset by integrating Faster-RCNN as the same baseline model our new method outperforms the peer method Gliding Vertex by 1.13% mAP50 (relative improvement 1.54%) and 2.46% mAP75 (relative improvement 5.91%) without any tricks.

Pose refinement is an interesting and practically relevant research direction. Pose refinement can be used to (1) obtain a more accurate pose estimate from an initial prior (e.g. from retrieval) (2) as pre-processing i.e. to provide a better starting point to a more expensive pose estimator (3) as post-processing of a more accurate localizer. Existing approaches focus on learning features / scene representations for the pose refinement task. This involves training an implicit scene representation or learning features while optimizing a camera pose-based loss. A natural question is whether training specific features / representations is truly necessary or whether similar results can be already achieved with more generic features. In this work we present a simple approach that combines pre-trained features with a particle filter and a renderable representation of the scene. Despite its simplicity it achieves state-of-the-art results demonstrating that one can easily build a pose refiner without the need for specific training. The code will be released upon acceptance.

Recognizing continuous changes offers valuable insights into past historical events supports current trend analysis and facilitates future planning. This knowledge is crucial for a variety of fields such as meteorology and agriculture environmental science urban planning and construction tourism and cultural preservation. Currently available datasets in the field of scene change understanding primarily concentrate on two main tasks: the detection of changed regions within a scene and the linguistic description of the change content. Existing datasets focus on recognizing discrete changes such as adding or deleting an object from two images and largely rely on artificially generated images. Consequently the existing change understanding methods primarily focus on identifying distinct object differences overlooking the importance of continuous gradual changes occurring over extended time intervals. To address the above issues we propose a novel benchmark dataset STVchrono targeting the localization and description of long-term continuous changes in real-world scenes. The dataset consists of 71900 photographs from Google Street View API taken over an 18-year span across 50 cities all over the world. Our STVchrono dataset is designed to support real-world continuous change recognition and description in both image pairs and extended image sequences while also enabling the segmentation of changed regions. We conduct experiments to evaluate state-of-the-art methods on continuous change description and segmentation as well as multimodal Large Language Models for describing changes. Our findings reveal that even the most advanced methods lag human performance emphasizing the need to adapt them to continuously changing real-world scenarios. We hope that our benchmark dataset will further facilitate the research of temporal change recognition in a dynamic world. The STVchrono dataset is available at STVchrono Dataset.

Vision-language models (VLMs) excel in zero-shot recognition but their performance varies greatly across different visual concepts. For example although CLIP achieves impressive accuracy on ImageNet (60-80%) its performance drops below 10% for more than ten concepts like night snake presumably due to their limited presence in the pretraining data. However measuring the frequency of concepts in VLMs' large-scale datasets is challenging. We address this by using large language models (LLMs) to count the number of pretraining texts that contain synonyms of these concepts. Our analysis confirms that popular datasets such as LAION exhibit a long-tailed concept distribution yielding biased performance in VLMs. We also find that downstream applications of VLMs including visual chatbots (e.g. GPT-4V) and text-to-image models (e.g. Stable Diffusion) often fail to recognize or generate images of rare concepts identified by our method. To mitigate the imbalanced performance of zero-shot VLMs we propose REtrieval-Augmented Learning (REAL). First instead of prompting VLMs using the original class names REAL uses their most frequent synonyms found in pretraining texts. This simple change already outperforms costly human-engineered and LLM-enriched prompts over nine benchmark datasets. Second REAL trains a linear classifier on a small yet balanced set of pretraining data retrieved using concept synonyms. REAL surpasses the previous zero-shot SOTA using 400x less storage and 10000x less training time!

Humans can infer 3D structure from 2D images of an object based on past experience and improve their 3D understanding as they see more images. Inspired by this behavior we introduce SAP3D a system for 3D reconstruction and novel view synthesis from an arbitrary number of unposed images. Given a few unposed images of an object we adapt a pre-trained view-conditioned diffusion model together with the camera poses of the images via test-time fine-tuning. The adapted diffusion model and the obtained camera poses are then utilized as instance-specific priors for 3D reconstruction and novel view synthesis. We show that as the number of input images increases the performance of our approach improves bridging the gap between optimization-based prior-less 3D reconstruction methods and single-image-to-3D diffusion-based methods. We demonstrate our system on real images as well as standard synthetic benchmarks. Our ablation studies confirm that this adaption behavior is key for more accurate 3D understanding.

Large-scale black-box models have become ubiquitous across numerous applications. Understanding the influence of individual training data sources on predictions made by these models is crucial for improving their trustworthiness. Current influence estimation techniques involve computing gradients for every training point or repeated training on different subsets. These approaches face obvious computational challenges when scaled up to large datasets and models. In this paper we introduce and explore the Mirrored Influence Hypothesis highlighting a reciprocal nature of influence between training and test data. Specifically it suggests that evaluating the influence of training data on test predictions can be reformulated as an equivalent yet inverse problem: assessing how the predictions for training samples would be altered if the model were trained on specific test samples. Through both empirical and theoretical validations we demonstrate the wide applicability of our hypothesis. Inspired by this we introduce a new method for estimating the influence of training data which requires calculating gradients for specific test samples paired with a forward pass for each training point. This approach can capitalize on the common asymmetry in scenarios where the number of test samples under concurrent examination is much smaller than the scale of the training dataset thus gaining a significant improvement in efficiency compared to existing approaches. We demonstrate the applicability of our method across a range of scenarios including data attribution in diffusion models data leakage detection analysis of memorization mislabeled data detection and tracing behavior in language models.

In the past few decades Japanese comics commonly referred to as Manga have transcended both cultural and linguistic boundaries to become a true worldwide sensation. Yet the inherent reliance on visual cues and illustration within manga renders it largely inaccessible to individuals with visual impairments. In this work we seek to address this substantial barrier with the aim of ensuring that manga can be appreciated and actively engaged by everyone. Specifically we tackle the problem of diarisation i.e. generating a transcription of who said what and when in a fully automatic way. To this end we make the following contributions: (1) we present a unified model Magi that is able to (a) detect panels text boxes and character boxes (b) cluster characters by identity (without knowing the number of clusters apriori) and (c) associate dialogues to their speakers; (2) we propose a novel approach that is able to sort the detected text boxes in their reading order and generate a dialogue transcript; (3) we annotate an evaluation benchmark for this task using publicly available [English] manga pages.

Recent advancements in large vision-language models enabled visual object detection in open-vocabulary scenarios where object classes are defined in free-text formats during inference. In this paper we aim to probe the state-of-the-art methods for open-vocabulary object detection to determine to what extent they understand fine-grained properties of objects and their parts. To this end we introduce an evaluation protocol based on dynamic vocabulary generation to test whether models detect discern and assign the correct fine-grained description to objects in the presence of hard-negative classes. We contribute with a benchmark suite of increasing difficulty and probing different properties like color pattern and material. We further enhance our investigation by evaluating several state-of-the-art open-vocabulary object detectors using the proposed protocol and find that most existing solutions which shine in standard open-vocabulary benchmarks struggle to accurately capture and distinguish finer object details. We conclude the paper by highlighting the limitations of current methodologies and exploring promising research directions to overcome the discovered drawbacks. Data and code are available at https://lorebianchi98.github.io/FG-OVD .