ECCV 2024 Papers

"Controlling the spectral norm of the Jacobian matrix, which is related to the convolution operation, has been shown to improve generalization, training stability and robustness in CNNs. Existing methods for computing the norm either tend to overestimate it or their performance may deteriorate quickly with increasing the input and kernel sizes. In this paper, we demonstrate that the tensor version of the spectral norm of a four-dimensional convolution kernel, up to a constant factor, serves as an upper bound for the spectral norm of the Jacobian matrix associated with the convolution operation. This new upper bound is independent of the input image resolution, differentiable and can be efficiently calculated during training. Through experiments, we demonstrate how this new bound can be used to improve the performance of convolutional architectures."

"In this paper, we investigate how to conduct transfer learning to adapt Stable Diffusion to downstream visual dense prediction tasks such as semantic segmentation and depth estimation. We focus on fine-tuning the Stable Diffusion model, which has demonstrated impressive abilities in modeling image details and high-level semantics. Through our experiments, we have three key insights. Firstly, we demonstrate that for dense prediction tasks, the denoiser of Stable Diffusion can serve as a stronger feature encoder compared to visual-language models pre-trained with contrastive training (e.g., CLIP). Secondly, we show that the quality of extracted features is influenced by the diffusion sampling step t, sampling layer, cross-attention map, model generation capacity, and textual input. Features from Stable Diffusion UNet’s upsampling layers and earlier denoising steps lead to more discriminative features for transfer learning to downstream tasks. Thirdly, we find that tuning Stable Diffusion to downstream tasks in a parameter-efficient way is feasible. We first extensively investigate currently popular parameter-efficient tuning methods. Then we search for the best protocol for effective tuning via reinforcement learning and achieve better tuning results with fewer tunable parameters."

"We combine concept-based neural networks with generative, flow-based classifiers into a novel, intrinsically explainable, exactly invertible approach to supervised learning. Prototypical neural networks, a type of concept-based neural network, represent an exciting way forward in realizing human-comprehensible machine learning without concept annotations, but a human-machine semantic gap continues to haunt current approaches. We find that reliance on indirect interpretation functions for prototypical explanations imposes a severe limit on prototypes’ informative power. From this, we posit that invertibly learning prototypes as distributions over the latent space provides more robust, expressive, and interpretable modeling. We propose one such model, called , by composing a normalizing flow with Gaussian mixture models. (1) sets a new state-of-the-art in joint generative and predictive modeling and (2) achieves predictive performance comparable to existing prototypical neural networks while enabling richer interpretation."

"Compositing an object into an image involves multiple non-trivial sub-tasks such as object placement and scaling, color/lighting harmonization, viewpoint/geometry adjustment, and shadow/reflection generation. Recent generative image compositing methods leverage diffusion models to handle multiple sub-tasks at once. However, existing models face limitations due to their reliance on masking the original object during training, which constrains their generation to the input mask. Furthermore, obtaining an accurate input mask specifying the location and scale of the object in a new image can be highly challenging. To overcome such limitations, we define a novel problem of unconstrained generative object compositing, i.e., the generation is not bounded by the mask, and train a diffusion-based model on a synthesized paired dataset. Our first-of-its-kind model is able to generate object effects such as shadows and reflections that go beyond the mask, enhancing image realism. Additionally, if an empty mask is provided, our model automatically places the object in diverse natural locations and scales, accelerating the compositing workflow. Our model outperforms existing object placement and compositing models in various quality metrics and user studies."

"Real-world autonomous driving (AD) like urban driving involves many corner cases. The lately released AD Benchmark CARLA Leaderboard v2 (a.k.a. CARLA v2) involves 39 new common events in the driving scene, providing a more quasi-realistic testbed compared to CARLA Leaderboard v1. It poses new challenges and so far no literature has reported any success on the new scenarios in v2. In this work, we take the initiative of directly training a neural planner and the hope is to handle the corner cases flexibly and effectively. To our best knowledge, we develop the first model-based RL method (named Think2Drive) for AD, with a compact latent world model to learn the transitions of the environment, and then it acts as a neural simulator to train the agent i.e. planner. It significantly boosts the training efficiency of RL thanks to the low dimensional state space and parallel computing of tensors in the latent world model. Think2Drive is able to run in an expert-level proficiency in CARLA v2 within 3 days of training on a single A6000 GPU, and to our best knowledge, so far there is no reported success (100% route completion) on CARLA v2. We also develop CornerCaseRepo, a benchmark that supports the evaluation of driving models by scenarios. We also propose a balanced metric to evaluate the performance by route completion, infraction number, and scenario density."

"Object placement is a task to insert a foreground object into a background scene at a suitable position and size. Existing methods mainly focus on extracting better visual features, while neglecting common sense about the objects and background. It leads to semantically unrealistic object positions. In this paper, we introduce Think Before Placement, a novel framework that effectively combines the implicit and explicit knowledge to generate placements that are both visually coherent and contextually appropriate. Specifically, we first adopt a large multi-modal model to generate a descriptive caption that identifies an appropriate position in the background for placing foreground object(Think ), then output proper position and size of the object (Place). The caption serves as an explicit semantic guidance for the subsequent placement of objects. Using this framework, we implement our model named CSENet, which outperforms baseline methods on the OPA dataset in extensive experiments. Further, we establish the OPAZ dataset to evaluate the zero-shot transfer capabilities of CSENet, where it also shows impressive performance across different foreground objects and scenes."

"Novel-view synthesis based on visible light has been extensively studied. In comparison to visible light imaging, thermal infrared imaging offers the advantage of all-weather imaging and strong penetration, providing increased possibilities for reconstruction in nighttime and adverse weather scenarios. However, thermal infrared imaging is influenced by physical characteristics such as atmospheric transmission effects and thermal conduction, hindering the precise reconstruction of intricate details in thermal infrared scenes, manifesting as issues of floaters and indistinct edge features in synthesized images. To address these limitations, this paper introduces a physics-induced 3D Gaussian splatting method named Thermal3D-GS. Thermal3D-GS begins by modeling atmospheric transmission effects and thermal conduction in three-dimensional media using neural networks. Additionally, a temperature consistency constraint is incorporated into the optimization objective to enhance the reconstruction accuracy of thermal infrared images. Furthermore, to validate the effectiveness of our method, the first large-scale benchmark dataset for this field named Thermal Infrared Novel-view Synthesis Dataset (TI-NSD) is created. This dataset comprises 20 authentic thermal infrared video scenes, covering indoor, outdoor, and UAV(Unmanned Aerial Vehicle) scenarios, totaling 6,664 frames of thermal infrared image data. Based on this dataset, this paper experimentally verifies the effectiveness of Thermal3D-GS. The results indicate that our method outperforms the baseline method with a 3.03 dB improvement in PSNR and significantly addresses the issues of floaters and indistinct edge features present in the baseline method. Our dataset and codebase will be released in redThermal3DGS."

"Inverse rendering of outdoor scenes from unconstrained image collections is a challenging task, particularly illumination/albedo ambiguities and occlusion of the illumination environment (shadowing) caused by geometry. However, there are many cues in an image that can aid in the disentanglement of geometry, albedo and shadows. Whilst sky is frequently masked out in state-of-the-art methods, we exploit the fact that any sky pixel provides a direct observation of distant lighting in the corresponding direction and, via a neural illumination prior, a statistical cue to derive the remaining illumination environment. The incorporation of our illumination prior is enabled by a novel ‘outside-in’ method for computing differentiable sky visibility based on a neural directional distance function. This is highly efficient and can be trained in parallel with the neural scene representation, allowing gradients from appearance loss to flow from shadows to influence the estimation of illumination and geometry. Our method estimates high-quality albedo, geometry, illumination and sky visibility, achieving state-of-the-art results on the NeRF-OSR relighting benchmark. Our code and models can be found at https://github.com/JADGardner/neusky."

"Knowledge distillation is an effective method for training lightweight vision models. However, acquiring teacher supervision for training samples is often costly, especially from large-scale models like vision transformers (ViTs). In this paper, we develop a simple framework to reduce the supervision cost of ViT distillation: masking out a fraction of input tokens given to the teacher. By masking input tokens, one can skip the computations associated with the masked tokens without requiring any change to teacher parameters or architecture. We find that masking patches with the lowest student attention scores is highly effective, saving up to 50% of teacher FLOPs without any drop in student accuracy, while other masking criterion leads to suboptimal efficiency gains. Through in-depth analyses, we reveal that the student-guided masking provides a good curriculum to the student, making teacher supervision easier to follow during the early stage and challenging in the later stage."

"In this work, we propose the use of Neural Radiance Fields () as a scene representation for visual localization. Recently, has been employed to enhance pose regression and scene coordinate regression models by augmenting the training database, providing auxiliary supervision through rendered images, or serving as an iterative refinement module. We extend its recognized advantages – its ability to provide a compact scene representation with realistic appearances and accurate geometry – by exploring the potential of ’s internal features in establishing precise 2D-3D matches for localization. To this end, we conduct a comprehensive examination of ’s implicit knowledge, acquired through view synthesis, for matching under various conditions. This includes exploring different matching network architectures, extracting encoder features at multiple layers, and varying training configurations. Significantly, we introduce , an advanced 2D-3D matching function that capitalizes on the internal knowledge of learned via view synthesis. Our evaluation of on standard localization benchmarks, within a structure-based pipeline, achieves competitive results for localization performance on Cambridge Landmarks. We will release all models and code."

"Text-to-image diffusion models allow users control over the content of generated images. Still, text-to-image generation occasionally leads to generation failure requiring users to generate dozens of images under the same text prompt before they obtain a satisfying result. We formulate the lottery ticket hypothesis in denoising: randomly initialized Gaussian noise images contain special pixel blocks (winning tickets) that naturally tend to be denoised into specific content independently. The generation failure in standard text-to-image synthesis is caused by the gap between optimal and actual spatial distribution of winning tickets in initial noisy images. To this end, we implement semantic-driven initial image construction creating initial noise from known winning tickets for each concept mentioned in the prompt. We conduct a series of experiments that verify the properties of winning tickets and demonstrate their generalizability across images and prompts. Our results show that aggregating winning tickets into the initial noise image effectively induce the model to generate the specified object at the corresponding location."

"Several benchmarks have concluded that our best vision-language models (, CLIP) are lacking in compositionality. Given an image, these benchmarks probe a model’s ability to identify its associated caption amongst a set of compositional distractors. In response, a surge of recent proposals show improvements by finetuning CLIP with distractors as hard negatives. Our investigations reveal that these improvements have, in fact, been overstated — because existing benchmarks do not probe whether finetuned models remain invariant to hard positives. By curating an evaluation dataset with 112, 382 hard negatives and hard positives, we uncover that including hard positives decreases CLIP’s performance by 12.9%, while humans perform effortlessly at 99%. CLIP finetuned with hard negatives results in an even larger decrease, up to 38.7%. With this finding, we then produce a 1,775,259 image-text training set with both hard negative and hard positive captions. By training with both, we see improvements on existing benchmarks while simultaneously improving performance on hard positives, indicating a more robust improvement in compositionality. Our work suggests the need for future research to rigorously test and improve CLIP’s understanding of semantic relationships between related “positive” concepts."

"Visual counterfactual explanation (CF) methods modify image concepts, , shape, to change a prediction to a predefined outcome while closely resembling the original query image. Unlike self-explainable models (SEMs) and heatmap techniques, they grant users the ability to examine hypothetical ”what-if” scenarios. Previous CF methods either entail post-hoc training, limiting the balance between transparency and CF quality, or demand optimization during inference. To bridge the gap between transparent SEMs and CF methods, we introduce the GdVAE, a self-explainable model based on a conditional variational autoencoder (CVAE), featuring a Gaussian discriminant analysis (GDA) classifier and integrated CF explanations. Full transparency is achieved through a generative classifier that leverages class-specific prototypes for the downstream task and a closed-form solution for CFs in the latent space. The consistency of CFs is improved by regularizing the latent space with the explainer function. Extensive comparisons with existing approaches affirm the effectiveness of our method in producing high-quality CF explanations while preserving transparency. Code and models are public."

"Large vision-language models (LVLMs), designed to interpret and respond to human instructions, occasionally generate hallucinated or harmful content due to inappropriate instructions. This study uses linear probing to shed light on the hidden knowledge at the output layers of LVLMs. We demonstrate that the logit distributions of the first tokens contain sufficient information to determine whether to respond to the instructions, including recognizing unanswerable visual questions, defending against jailbreaking attacks, and identifying deceptive questions. Such hidden knowledge is gradually lost in logits of subsequent tokens during response generation. Then, we illustrate a simple decoding strategy at the generation of the first token, effectively improving the generated content. In experiments, we find a few interesting insights: First, the CLIP model already contains a strong signal for solving these tasks, which indicates potential bias in the existing datasets. Second, we observe performance improvement by utilizing the first logit distributions on three additional tasks, including indicating uncertainty in math solving, mitigating hallucination, and image classification. Last, with the same training data, simply finetuning LVLMs improves models’ performance but is still inferior to linear probing on these tasks. 1 1 Our code is available at https://github.com/Qinyu-Allen-Zhao/LVLM-LP."

"In spite of recent advancements in text-to-image generation, limitations persist in handling complex and imaginative prompts due to the restricted diversity and complexity of training data. This work explores how diffusion models can generate images from prompts requiring artistic creativity or specialized knowledge. We introduce the Realistic-Fantasy Benchmark (RFBench), a novel evaluation framework blending realistic and fantastical scenarios. To address these challenges, we propose the Realistic-Fantasy Network (RFNet), a training-free approach integrating diffusion models with LLMs. Extensive human evaluations and GPT-based compositional assessments demonstrate our approach’s superiority over state-of-the-art methods. Our code and dataset is available at https: //leo81005.github.io/Reality-and-Fantasy/."

"Despite the recent success of domain generalization in medical image segmentation, voxel-wise annotation for all source domains remains a huge burden. Semi-supervised domain generalization has been proposed very recently to combat this challenge by leveraging limited labeled data along with abundant unlabeled data collected from multiple medical institutions, depending on precisely harnessing unlabeled data while improving generalization simultaneously. In this work, we observe that domain shifts between medical institutions cause disparate feature statistics, which significantly deteriorates pseudo-label quality due to an unexpected normalization process. Nevertheless, this phenomenon could be exploited to facilitate unseen domain generalization. Therefore, we propose 1) multiple statistics-individual branches to mitigate the impact of domain shifts for reliable pseudo-labels and 2) one statistics-aggregated branch for domain-invariant feature learning. Furthermore, to simulate unseen domains with statistics difference, we approach this from two aspects, , a perturbation with histogram matching at image level and a random batch normalization selection strategy at feature level, producing diverse statistics to expand the training distribution. Evaluation results on three medical image datasets demonstrate the effectiveness of our method compared with recent SOTA methods. The code is available at https: //github.com/qiumuyang/SIAB."

"We present the All-Seeing Project V2: a new model and dataset designed for understanding object relations in images. Specifically, we propose the All-Seeing Model V2 () that integrates the formulation of text generation, object localization, and relation comprehension into a relation conversation (ReC) task. Leveraging this unified task, our model excels not only in perceiving and recognizing all objects within the image but also in grasping the intricate relation graph between them, diminishing the relation hallucination often encountered by Multi-modal Large Language Models (MLLMs). To facilitate training and evaluation of MLLMs in relation understanding, we created the first high-quality ReC dataset () which is aligned with the format of standard instruction tuning data. In addition, we design a new benchmark, termed Circular-based Relation Probing Evaluation () for comprehensively evaluating the relation comprehension capabilities of MLLMs. Notably, our achieves an overall accuracy of 64.50 on this relation-aware benchmark, surpassing the 55.63 of LLaVA-1.5 by a large margin. We hope that our work can inspire more future research and contribute to the evolution towards artificial general intelligence. Our project is released at https:// github.com/OpenGVLab/all-seeing."

"3D Gaussian splatting, emerging as a groundbreaking approach, has drawn increasing attention for its capabilities of high-fidelity reconstruction and real-time rendering. However, it couples the appearance and geometry of the scene within the Gaussian attributes, which hinders the flexibility of editing operations, such as texture swapping. To address this issue, we propose a novel approach, namely Texture-GS, to disentangle the appearance from the geometry by representing it as a 2D texture mapped onto the 3D surface, thereby facilitating appearance editing. Technically, the disentanglement is achieved by our proposed texture mapping module, which consists of a UV mapping MLP to learn the UV coordinates for the 3D Gaussian centers, a local Taylor expansion of the MLP to efficiently approximate the UV coordinates for the ray-Gaussian intersections, and a learnable texture to capture the fine-grained appearance. Extensive experiments on the DTU dataset demonstrate that our method not only facilitates high-fidelity appearance editing but also achieves real-time rendering on consumer-level devices, a single RTX 2080 Ti GPU."

"Text-to-image generation plays a pivotal role in computer vision and natural language processing by translating textual descriptions into visual representations. However, understanding complex relations in detailed text prompts filled with rich relational content remains a significant challenge. To address this, we introduce a novel task: Logic-Rich Text-to-Image generation. Unlike conventional image generation tasks that rely on short and structurally simple natural language inputs, our task focuses on intricate text inputs abundant in relational information. To tackle these complexities, we collect the Textual-Visual Logic dataset, designed to evaluate the performance of text-to-image generation models across diverse and complex scenarios. Furthermore, we propose a baseline model as a benchmark for this task. Our model comprises three key components: a relation understanding module, a multimodality fusion module, and a negative pair discriminator. These components enhance the model’s ability to handle disturbances in informative tokens and prioritize relational elements during image generation. https:// github.com/IntelLabs/Textual-Visual-Logic-Challenge"