Semantic image synthesis, i.e., generating images from user-provided semantic label maps, is an important conditional image generation task as it allows to control both the content as well as the spatial layout of generated images. Although diffusion models have pushed the state of the art in generative image modeling, the iterative nature of their inference process makes them computationally demanding. Other approaches such as GANs are more efficient as they only need a single feed-forward pass for generation, but the image quality tends to suffer on large and diverse datasets. In this work, we propose a new class of GAN discriminators for semantic image synthesis that generates highly realistic images by exploiting feature backbone networks pre-trained for tasks such as image classification. We also introduce a new generator architecture with better context modeling and using cross-attention to inject noise into latent variables, leading to more diverse generated images. Our model, which we dub DP-SIMS, achieves state-of-the-art results in terms of image quality and consistency with the input label maps on ADE-20K, COCO-Stuff, and Cityscapes, surpassing recent diffusion models while requiring two orders of magnitude less compute for inference.
Recent studies on generalizing CLIP for monocular depth estimation reveal that CLIP pre-trained on web-crawled data is inefficient for deriving proper similarities between image patches and depth-related prompts. In this paper, we adapt CLIP for meaningful quality of monocular depth estimation with dense prediction, without fine-tuning its original vision-language alignment. By jointly training a compact deconvolutional decoder with a tiny learnable embedding matrix named mirror, as a static prompt for its text encoder, CLIP is enabled to understand depth. With this approach, our model exhibits impressive performance matching several previous state-of-the-art vision-only models on the NYU Depth v2 and KITTI datasets, outperforming every CLIP-based depth estimation model with a large margin. Experiments on temporal depth consistency and spatial continuity demonstrate that the prior knowledge of CLIP can be effectively refined by our proposed framework. Furthermore, an ablation study on mirror proves that the resulting model estimates depth utilizing knowledge not only from the image encoder but also text encoder despite not being given any prompt written in a human way. This research demonstrates that through minimal adjustments, the prior knowledge of vision-language foundation models, such as CLIP, can be generalized even to domains where learning during pretraining is challenging. We facilitate future works focused on methods to adjust suboptimal prior knowledge of vision-language models using non-human language prompts, achieving performance on par with task-specific state-of-the-art methodologies.
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.
Digital pixel sensor (DPS) has evolved as a pivotal component in modern imaging systems and has the potential to revolutionize various fields such as medical imaging, astronomy, surveillance, IoT devices, etc. Compared to analog pixel sensors, the DPS offers high speed and good image quality. However, the introduced intrinsic complexity within each pixel, primarily attributed to the accommodation of the ADC circuit, engenders a substantial increase in the pixel pitch. Unfortunately, such a pronounced escalation in pixel pitch drastically undermines the feasibility of achieving high-density integration, which is an obstacle that significantly narrows down the field of potential applications. Nonetheless, designing compact conversion circuits along with strategic integration of 3D architectural paradigms can be a potential remedy to the prevailing situation. This review article presents a comprehensive overview of the vast area of DPS technology. The operating principles, advantages, and challenges of different types of DPS circuits have been analyzed. We categorize the schemes into several categories based on ADC operation. A comparative study based on different performance metrics has also been showcased for a well-rounded understanding.
3D content creation has achieved significant progress in terms of both quality and speed. Although current feed-forward models can produce 3D objects in seconds, their resolution is constrained by the intensive computation required during training. In this paper, we introduce Large Multi-View Gaussian Model (LGM), a novel framework designed to generate high-resolution 3D models from text prompts or single-view images. Our key insights are two-fold: 1) 3D Representation: We propose multi-view Gaussian features as an efficient yet powerful representation, which can then be fused together for differentiable rendering. 2) 3D Backbone: We present an asymmetric U-Net as a high-throughput backbone operating on multi-view images, which can be produced from text or single-view image input by leveraging multi-view diffusion models. Extensive experiments demonstrate the high fidelity and efficiency of our approach. Notably, we maintain the fast speed to generate 3D objects within 5 seconds while boosting the training resolution to 512, thereby achieving high-resolution 3D content generation.
Semantic mapping based on the supervised object detectors is sensitive to image distribution. In real-world environments, the object detection and segmentation performance can lead to a major drop, preventing the use of semantic mapping in a wider domain. On the other hand, the development of vision-language foundation models demonstrates a strong zero-shot transferability across data distribution. It provides an opportunity to construct generalizable instance-aware semantic maps. Hence, this work explores how to boost instance-aware semantic mapping from object detection generated from foundation models. We propose a probabilistic label fusion method to predict close-set semantic classes from open-set label measurements. An instance refinement module merges the over-segmented instances caused by inconsistent segmentation. We integrate all the modules into a unified semantic mapping system. Reading a sequence of RGB-D input, our work incrementally reconstructs an instance-aware semantic map. We evaluate the zero-shot performance of our method in ScanNet and SceneNN datasets. Our method achieves 40.3 mean average precision (mAP) on the ScanNet semantic instance segmentation task. It outperforms the traditional semantic mapping method significantly.
Human de-occlusion, which aims to infer the appearance of invisible human parts from an occluded image, has great value in many human-related tasks, such as person re-id, and intention inference. To address this task, this paper proposes a dynamic mask-aware transformer (DMAT), which dynamically augments information from human regions and weakens that from occlusion. First, to enhance token representation, we design an expanded convolution head with enlarged kernels, which captures more local valid context and mitigates the influence of surrounding occlusion. To concentrate on the visible human parts, we propose a novel dynamic multi-head human-mask guided attention mechanism through integrating multiple masks, which can prevent the de-occluded regions from assimilating to the background. Besides, a region upsampling strategy is utilized to alleviate the impact of occlusion on interpolated images. During model learning, an amodal loss is developed to further emphasize the recovery effect of human regions, which also refines the model's convergence. Extensive experiments on the AHP dataset demonstrate its superior performance compared to recent state-of-the-art methods.
Deep neural networks are exploited using natural adversarial samples, which have no impact on human perception but are misclassified. Current approaches often rely on the white-box nature of deep neural networks to generate these adversarial samples or alter the distribution of adversarial samples compared to training distribution. To alleviate the limitations of current approaches, we propose EvoSeed, a novel evolutionary strategy-based search algorithmic framework to generate natural adversarial samples. Our EvoSeed framework uses auxiliary Diffusion and Classifier models to operate in a model-agnostic black-box setting. We employ CMA-ES to optimize the search for an adversarial seed vector, which, when processed by the Conditional Diffusion Model, results in an unrestricted natural adversarial sample misclassified by the Classifier Model. Experiments show that generated adversarial images are of high image quality and are transferable to different classifiers. Our approach demonstrates promise in enhancing the quality of adversarial samples using evolutionary algorithms. We hope our research opens new avenues to enhance the robustness of deep neural networks in real-world scenarios. Project Website can be accessed at \url{https://shashankkotyan.github.io/EvoSeed}.
In this work we study the enhancement of Low Rank Adaptation (LoRA) fine-tuning procedure by introducing a Riemannian preconditioner in its optimization step. Specifically, we introduce an $r\times r$ preconditioner in each gradient step where $r$ is the LoRA rank. This preconditioner requires a small change to existing optimizer code and creates virtually minuscule storage and runtime overhead. Our experimental results with both large language models and text-to-image diffusion models show that with our preconditioner, the convergence and reliability of SGD and AdamW can be significantly enhanced. Moreover, the training process becomes much more robust to hyperparameter choices such as learning rate. Theoretically, we show that fine-tuning a two-layer ReLU network in the convex paramaterization with our preconditioner has convergence rate independent of condition number of the data matrix. This new Riemannian preconditioner, previously explored in classic low-rank matrix recovery, is introduced to deep learning tasks for the first time in our work. We release our code at https://github.com/pilancilab/Riemannian_Preconditioned_LoRA.
Visualizing CNN is an important part in building trust and explaining model's prediction. Methods like CAM and GradCAM have been really successful in localizing area of the image responsible for the output but are only limited to classification models. In this paper, we present a new method EmbeddingCAM, which generalizes the Grad-CAM for embedding networks. We show that for classification networks, EmbeddingCAM reduces to GradCAM. We show the effectiveness of our method on CUB-200-2011 dataset and also present quantitative and qualitative analysis on the dataset.