The emergence of large multimodal models has unlocked remarkable potential in AI, particularly in pathology. However, the lack of specialized, high-quality benchmark impeded their development and precise evaluation. To address this, we introduce PathMMU, the largest and highest-quality expert-validated pathology benchmark for LMMs. It comprises 33,573 multimodal multi-choice questions and 21,599 images from various sources, and an explanation for the correct answer accompanies each question. The construction of PathMMU capitalizes on the robust capabilities of GPT-4V, utilizing approximately 30,000 gathered image-caption pairs to generate Q\&As. Significantly, to maximize PathMMU's authority, we invite six pathologists to scrutinize each question under strict standards in PathMMU's validation and test sets, while simultaneously setting an expert-level performance benchmark for PathMMU. We conduct extensive evaluations, including zero-shot assessments of 14 open-sourced and three closed-sourced LMMs and their robustness to image corruption. We also fine-tune representative LMMs to assess their adaptability to PathMMU. The empirical findings indicate that advanced LMMs struggle with the challenging PathMMU benchmark, with the top-performing LMM, GPT-4V, achieving only a 51.7\% zero-shot performance, significantly lower than the 71.4\% demonstrated by human pathologists. After fine-tuning, even open-sourced LMMs can surpass GPT-4V with a performance of over 60\%, but still fall short of the expertise shown by pathologists. We hope that the PathMMU will offer valuable insights and foster the development of more specialized, next-generation LLMs for pathology.
Given the broad application of infrared technology across diverse fields, there is an increasing emphasis on investigating super-resolution techniques for infrared images within the realm of deep learning. Despite the impressive results of current Transformer-based methods in image super-resolution tasks, their reliance on the self-attentive mechanism intrinsic to the Transformer architecture results in images being treated as one-dimensional sequences, thereby neglecting their inherent two-dimensional structure. Moreover, infrared images exhibit a uniform pixel distribution and a limited gradient range, posing challenges for the model to capture effective feature information. Consequently, we suggest a potent Transformer model, termed Large Kernel Transformer (LKFormer), to address this issue. Specifically, we have designed a Large Kernel Residual Attention (LKRA) module with linear complexity. This mainly employs depth-wise convolution with large kernels to execute non-local feature modeling, thereby substituting the standard self-attentive layer. Additionally, we have devised a novel feed-forward network structure called Gated-Pixel Feed-Forward Network (GPFN) to augment the LKFormer's capacity to manage the information flow within the network. Comprehensive experimental results reveal that our method surpasses the most advanced techniques available, using fewer parameters and yielding considerably superior performance.The source code will be available at https://github.com/sad192/large-kernel-Transformer.
Reconstructing an object from photos and placing it virtually in a new environment goes beyond the standard novel view synthesis task as the appearance of the object has to not only adapt to the novel viewpoint but also to the new lighting conditions and yet evaluations of inverse rendering methods rely on novel view synthesis data or simplistic synthetic datasets for quantitative analysis. This work presents a real-world dataset for measuring the reconstruction and rendering of objects for relighting. To this end, we capture the environment lighting and ground truth images of the same objects in multiple environments allowing to reconstruct the objects from images taken in one environment and quantify the quality of the rendered views for the unseen lighting environments. Further, we introduce a simple baseline composed of off-the-shelf methods and test several state-of-the-art methods on the relighting task and show that novel view synthesis is not a reliable proxy to measure performance. Code and dataset are available at https://github.com/isl-org/objects-with-lighting .
Fusion of a panchromatic (PAN) image and corresponding multispectral (MS) image is also known as pansharpening, which aims to combine abundant spatial details of PAN and spectral information of MS. Due to the absence of high-resolution MS images, available deep-learning-based methods usually follow the paradigm of training at reduced resolution and testing at both reduced and full resolution. When taking original MS and PAN images as inputs, they always obtain sub-optimal results due to the scale variation. In this paper, we propose to explore the self-supervised representation of pansharpening by designing a cross-predictive diffusion model, named CrossDiff. It has two-stage training. In the first stage, we introduce a cross-predictive pretext task to pre-train the UNet structure based on conditional DDPM, while in the second stage, the encoders of the UNets are frozen to directly extract spatial and spectral features from PAN and MS, and only the fusion head is trained to adapt for pansharpening task. Extensive experiments show the effectiveness and superiority of the proposed model compared with state-of-the-art supervised and unsupervised methods. Besides, the cross-sensor experiments also verify the generalization ability of proposed self-supervised representation learners for other satellite's datasets. We will release our code for reproducibility.
Despite recent advances in text-to-3D generative methods, there is a notable absence of reliable evaluation metrics. Existing metrics usually focus on a single criterion each, such as how well the asset aligned with the input text. These metrics lack the flexibility to generalize to different evaluation criteria and might not align well with human preferences. Conducting user preference studies is an alternative that offers both adaptability and human-aligned results. User studies, however, can be very expensive to scale. This paper presents an automatic, versatile, and human-aligned evaluation metric for text-to-3D generative models. To this end, we first develop a prompt generator using GPT-4V to generate evaluating prompts, which serve as input to compare text-to-3D models. We further design a method instructing GPT-4V to compare two 3D assets according to user-defined criteria. Finally, we use these pairwise comparison results to assign these models Elo ratings. Experimental results suggest our metric strongly align with human preference across different evaluation criteria.
Multimodal information extraction (MIE) gains significant attention as the popularity of multimedia content increases. However, current MIE methods often resort to using task-specific model structures, which results in limited generalizability across tasks and underutilizes shared knowledge across MIE tasks. To address these issues, we propose UMIE, a unified multimodal information extractor to unify three MIE tasks as a generation problem using instruction tuning, being able to effectively extract both textual and visual mentions. Extensive experiments show that our single UMIE outperforms various state-of-the-art (SoTA) methods across six MIE datasets on three tasks. Furthermore, in-depth analysis demonstrates UMIE's strong generalization in the zero-shot setting, robustness to instruction variants, and interpretability. Our research serves as an initial step towards a unified MIE model and initiates the exploration into both instruction tuning and large language models within the MIE domain. Our code, data, and model are available at https://github.com/ZUCC-AI/UMIE
In the realm of large language models (LLMs), enhancing instruction-following capability often involves curating expansive training data. This is achieved through two primary schemes: i) Scaling-Inputs: Amplifying (input, output) pairs per task instruction, aiming for better instruction adherence. ii) Scaling Input-Free Tasks: Enlarging tasks, each composed of an (instruction, output) pair (without requiring a separate input anymore). However, LLMs under Scaling-Inputs tend to be overly sensitive to inputs, leading to misinterpretation or non-compliance with instructions. Conversely, Scaling Input-Free Tasks demands a substantial number of tasks but is less effective in instruction following when dealing with instances in Scaling-Inputs. This work introduces MUFFIN, a new scheme of instruction-following dataset curation. Specifically, we automatically Scale Tasks per Input by diversifying these tasks with various input facets. Experimental results across four zero-shot benchmarks, spanning both Scaling-Inputs and Scaling Input-Free Tasks schemes, reveal that LLMs, at various scales, trained on MUFFIN generally demonstrate superior instruction-following capabilities compared to those trained on the two aforementioned schemes.
AI applications are becoming increasingly visible to the general public. There is a notable gap between the theoretical assumptions researchers make about computer vision models and the reality those models face when deployed in the real world. One of the critical reasons for this gap is a challenging problem known as distribution shift. Distribution shifts tend to vary with complexity of the data, dataset size, and application type. In our paper, we discuss the identification of such a prominent gap, exploring the concept of distribution shift and its critical significance. We provide an in-depth overview of various types of distribution shifts, elucidate their distinctions, and explore techniques within the realm of the data-centric domain employed to address them. Distribution shifts can occur during every phase of the machine learning pipeline, from the data collection stage to the stage of training a machine learning model to the stage of final model deployment. As a result, it raises concerns about the overall robustness of the machine learning techniques for computer vision applications that are deployed publicly for consumers. Different deep learning models each tailored for specific type of data and tasks, architectural pipelines; highlighting how variations in data preprocessing and feature extraction can impact robustness., data augmentation strategies (e.g. geometric, synthetic and learning-based); demonstrating their role in enhancing model generalization, and training mechanisms (e.g. transfer learning, zero-shot) fall under the umbrella of data-centric methods. Each of these components form an integral part of the neural-network we analyze contributing uniquely to strengthening model robustness against distribution shifts. We compare and contrast numerous AI models that are built for mitigating shifts in hidden stratification and spurious correlations, ...
Molecular docking is a key computational tool utilized to predict the binding conformations of small molecules to protein targets, which is fundamental in the design of novel drugs. Despite recent advancements in geometric deep learning-based approaches leading to improvements in blind docking efficiency, these methods have encountered notable challenges, such as limited generalization performance on unseen proteins, the inability to concurrently address the settings of blind docking and site-specific docking, and the frequent occurrence of physical implausibilities such as inter-molecular steric clash. In this study, we introduce DeltaDock, a robust and versatile framework designed for efficient molecular docking to overcome these challenges. DeltaDock operates in a two-step process: rapid initial complex structures sampling followed by multi-scale iterative refinement of the initial structures. In the initial stage, to sample accurate structures with high efficiency, we develop a ligand-dependent binding site prediction model founded on large protein models and graph neural networks. This model is then paired with GPU-accelerated sampling algorithms. The sampled structures are updated using a multi-scale iterative refinement module that captures both protein-ligand atom-atom interactions and residue-atom interactions in the following stage. Distinct from previous geometric deep learning methods that are conditioned on the blind docking setting, DeltaDock demonstrates superior performance in both blind docking and site-specific docking settings. Comprehensive experimental results reveal that DeltaDock consistently surpasses baseline methods in terms of docking accuracy. Furthermore, it displays remarkable generalization capabilities and proficiency for predicting physically valid structures, thereby attesting to its robustness and reliability in various scenarios.
We introduce MMMU: a new benchmark designed to evaluate multimodal models on massive multi-discipline tasks demanding college-level subject knowledge and deliberate reasoning. MMMU includes 11.5K meticulously collected multimodal questions from college exams, quizzes, and textbooks, covering six core disciplines: Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, and Tech & Engineering. These questions span 30 subjects and 183 subfields, comprising 30 highly heterogeneous image types, such as charts, diagrams, maps, tables, music sheets, and chemical structures. Unlike existing benchmarks, MMMU focuses on advanced perception and reasoning with domain-specific knowledge, challenging models to perform tasks akin to those faced by experts. Our evaluation of 14 open-source LMMs and the proprietary GPT-4V(ision) highlights the substantial challenges posed by MMMU. Even the advanced GPT-4V only achieves a 56% accuracy, indicating significant room for improvement. We believe MMMU will stimulate the community to build next-generation multimodal foundation models towards expert artificial general intelligence.