The rise of portable Lidar instruments, including their adoption in smartphones, opens the door to novel computational imaging techniques. Being an active sensing instrument, Lidar can provide complementary data to passive optical sensors, particularly in situations like low-light imaging where motion blur can affect photos. In this paper, we study if the depth information provided by mobile Lidar sensors is useful for the task of image deblurring and how to integrate it with a general approach that transforms any state-of-the-art neural deblurring model into a depth-aware one. To achieve this, we developed a universal adapter structure that efficiently preprocesses the depth information to modulate image features with depth features. Additionally, we applied a continual learning strategy to pretrained encoder-decoder models, enabling them to incorporate depth information as an additional input with minimal extra data requirements. We demonstrate that utilizing true depth information can significantly boost the effectiveness of deblurring algorithms, as validated on a dataset with real-world depth data captured by a smartphone Lidar.
There has been immense progress recently in the visual quality of Stable Diffusion-based Super Resolution (SD-SR). However, deploying large diffusion models on computationally restricted devices such as mobile phones remains impractical due to the large model size and high latency. This is compounded for SR as it often operates at high res (e.g. 4Kx3K). In this work, we introduce Edge-SD-SR, the first parameter efficient and low latency diffusion model for image super-resolution. Edge-SD-SR consists of ~169M parameters, including UNet, encoder and decoder, and has a complexity of only ~142 GFLOPs. To maintain a high visual quality on such low compute budget, we introduce a number of training strategies: (i) A novel conditioning mechanism on the low resolution input, coined bidirectional conditioning, which tailors the SD model for the SR task. (ii) Joint training of the UNet and encoder, while decoupling the encodings of the HR and LR images and using a dedicated schedule. (iii) Finetuning the decoder using the UNet's output to directly tailor the decoder to the latents obtained at inference time. Edge-SD-SR runs efficiently on device, e.g. it can upscale a 128x128 patch to 512x512 in 38 msec while running on a Samsung S24 DSP, and of a 512x512 to 2048x2048 (requiring 25 model evaluations) in just ~1.1 sec. Furthermore, we show that Edge-SD-SR matches or even outperforms state-of-the-art SR approaches on the most established SR benchmarks.
Knowledge Distillation (KD) is a promising approach for unsupervised Anomaly Detection (AD). However, the student network's over-generalization often diminishes the crucial representation differences between teacher and student in anomalous regions, leading to detection failures. To addresses this problem, the widely accepted Reverse Distillation (RD) paradigm designs the asymmetry teacher and student, using an encoder as teacher and a decoder as student. Yet, the design of RD does not ensure that the teacher encoder effectively distinguishes between normal and abnormal features or that the student decoder generates anomaly-free features. Additionally, the absence of skip connections results in a loss of fine details during feature reconstruction. To address these issues, we propose RD with Expert, which introduces a novel Expert-Teacher-Student network for simultaneous distillation of both the teacher encoder and student decoder. The added expert network enhances the student's ability to generate normal features and optimizes the teacher's differentiation between normal and abnormal features, reducing missed detections. Additionally, Guided Information Injection is designed to filter and transfer features from teacher to student, improving detail reconstruction and minimizing false positives. Experiments on several benchmarks prove that our method outperforms existing unsupervised AD methods under RD paradigm, fully unlocking RD's potential.
Recent sparse multi-view scene reconstruction advances like DUSt3R and MASt3R no longer require camera calibration and camera pose estimation. However, they only process a pair of views at a time to infer pixel-aligned pointmaps. When dealing with more than two views, a combinatorial number of error prone pairwise reconstructions are usually followed by an expensive global optimization, which often fails to rectify the pairwise reconstruction errors. To handle more views, reduce errors, and improve inference time, we propose the fast single-stage feed-forward network MV-DUSt3R. At its core are multi-view decoder blocks which exchange information across any number of views while considering one reference view. To make our method robust to reference view selection, we further propose MV-DUSt3R+, which employs cross-reference-view blocks to fuse information across different reference view choices. To further enable novel view synthesis, we extend both by adding and jointly training Gaussian splatting heads. Experiments on multi-view stereo reconstruction, multi-view pose estimation, and novel view synthesis confirm that our methods improve significantly upon prior art. Code will be released.
The primary objective of Optical Chemical Structure Recognition is to identify chemical structure images into corresponding markup sequences. However, the complex two-dimensional structures of molecules, particularly those with rings and multiple branches, present significant challenges for current end-to-end methods to learn one-dimensional markup directly. To overcome this limitation, we propose a novel Ring-Free Language (RFL), which utilizes a divide-and-conquer strategy to describe chemical structures in a hierarchical form. RFL allows complex molecular structures to be decomposed into multiple parts, ensuring both uniqueness and conciseness while enhancing readability. This approach significantly reduces the learning difficulty for recognition models. Leveraging RFL, we propose a universal Molecular Skeleton Decoder (MSD), which comprises a skeleton generation module that progressively predicts the molecular skeleton and individual rings, along with a branch classification module for predicting branch information. Experimental results demonstrate that the proposed RFL and MSD can be applied to various mainstream methods, achieving superior performance compared to state-of-the-art approaches in both printed and handwritten scenarios. The code is available at https://github.com/JingMog/RFL-MSD.
In this work, we compare the domain-specific translation performance of open-source autoregressive decoder-only large language models (LLMs) with task-oriented machine translation (MT) models. Our experiments focus on the medical domain and cover four language directions with varied resource availability: English-to-French, English-to-Portuguese, English-to-Swahili, and Swahili-to-English. Despite recent advancements, LLMs demonstrate a significant quality gap in specialized translation compared to multilingual encoder-decoder MT models such as NLLB-200. Our results indicate that NLLB-200 3.3B outperforms all evaluated LLMs in the 7-8B parameter range across three out of the four language directions. While fine-tuning improves the performance of LLMs such as Mistral and Llama, these models still underperform compared to fine-tuned NLLB-200 3.3B models. Our findings highlight the ongoing need for specialized MT models to achieve high-quality domain-specific translation, especially in medium-resource and low-resource settings. Moreover, the superior performance of larger LLMs over their 8B variants suggests potential value in pre-training domain-specific medium-sized language models, employing targeted data selection and knowledge distillation approaches to enhance both quality and efficiency in specialized translation tasks.
Large language models with a transformer-based encoder/decoder architecture, such as T5, have become standard platforms for supervised tasks. To bring these technologies to the clinical domain, recent work has trained new or adapted existing models to clinical data. However, the evaluation of these clinical T5 models and comparison to other models has been limited. Are the clinical T5 models better choices than FLAN-tuned generic T5 models? Do they generalize better to new clinical domains that differ from the training sets? We comprehensively evaluate these models across several clinical tasks and domains. We find that clinical T5 models provide marginal improvements over existing models, and perform worse when evaluated on different domains. Our results inform future choices in developing clinical LLMs.
The Segment Anything Model (SAM), originally built on a 2D Vision Transformer (ViT), excels at capturing global patterns in 2D natural images but struggles with 3D medical imaging modalities like CT and MRI. These modalities require capturing spatial information in volumetric space for tasks such as organ segmentation and tumor quantification. To address this challenge, we introduce RefSAM3D, which adapts SAM for 3D medical imaging by incorporating a 3D image adapter and cross-modal reference prompt generation. Our approach modifies the visual encoder to handle 3D inputs and enhances the mask decoder for direct 3D mask generation. We also integrate textual prompts to improve segmentation accuracy and consistency in complex anatomical scenarios. By employing a hierarchical attention mechanism, our model effectively captures and integrates information across different scales. Extensive evaluations on multiple medical imaging datasets demonstrate the superior performance of RefSAM3D over state-of-the-art methods. Our contributions advance the application of SAM in accurately segmenting complex anatomical structures in medical imaging.
In recent years, advanced U-like networks have demonstrated remarkable performance in medical image segmentation tasks. However, their drawbacks, including excessive parameters, high computational complexity, and slow inference speed, pose challenges for practical implementation in scenarios with limited computational resources. Existing lightweight U-like networks have alleviated some of these problems, but they often have pre-designed structures and consist of inseparable modules, limiting their application scenarios. In this paper, we propose three plug-and-play decoders by employing different discretization methods of the neural memory Ordinary Differential Equations (nmODEs). These decoders integrate features at various levels of abstraction by processing information from skip connections and performing numerical operations on upward path. Through experiments on the PH2, ISIC2017, and ISIC2018 datasets, we embed these decoders into different U-like networks, demonstrating their effectiveness in significantly reducing the number of parameters and FLOPs while maintaining performance. In summary, the proposed discretized nmODEs decoders are capable of reducing the number of parameters by about 20% ~ 50% and FLOPs by up to 74%, while possessing the potential to adapt to all U-like networks. Our code is available at https://github.com/nayutayuki/Lightweight-nmODE-Decoders-For-U-like-networks.
Brain tumor segmentation models have aided diagnosis in recent years. However, they face MRI complexity and variability challenges, including irregular shapes and unclear boundaries, leading to noise, misclassification, and incomplete segmentation, thereby limiting accuracy. To address these issues, we adhere to an outstanding Convolutional Neural Networks (CNNs) design paradigm and propose a novel network named A4-Unet. In A4-Unet, Deformable Large Kernel Attention (DLKA) is incorporated in the encoder, allowing for improved capture of multi-scale tumors. Swin Spatial Pyramid Pooling (SSPP) with cross-channel attention is employed in a bottleneck further to study long-distance dependencies within images and channel relationships. To enhance accuracy, a Combined Attention Module (CAM) with Discrete Cosine Transform (DCT) orthogonality for channel weighting and convolutional element-wise multiplication is introduced for spatial weighting in the decoder. Attention gates (AG) are added in the skip connection to highlight the foreground while suppressing irrelevant background information. The proposed network is evaluated on three authoritative MRI brain tumor benchmarks and a proprietary dataset, and it achieves a 94.4% Dice score on the BraTS 2020 dataset, thereby establishing multiple new state-of-the-art benchmarks. The code is available here: https://github.com/WendyWAAAAANG/A4-Unet.
The analysis of retinal images for the diagnosis of various diseases is one of the emerging areas of research. Recently, the research direction has been inclined towards investigating several changes in retinal blood vessels in subjects with many neurological disorders, including dementia. This research focuses on detecting diseases early by improving the performance of models for segmentation of retinal vessels with fewer parameters, which reduces computational costs and supports faster processing. This paper presents a novel lightweight encoder-decoder model that segments retinal vessels to improve the efficiency of disease detection. It incorporates multi-scale convolutional blocks in the encoder to accurately identify vessels of various sizes and thicknesses. The bottleneck of the model integrates the Focal Modulation Attention and Spatial Feature Refinement Blocks to refine and enhance essential features for efficient segmentation. The decoder upsamples features and integrates them with the corresponding feature in the encoder using skip connections and the spatial feature refinement block at every upsampling stage to enhance feature representation at various scales. The estimated computation complexity of our proposed model is around 29.60 GFLOP with 0.71 million parameters and 2.74 MB of memory size, and it is evaluated using public datasets, that is, DRIVE, CHASE\_DB, and STARE. It outperforms existing models with dice scores of 86.44\%, 84.22\%, and 87.88\%, respectively.
Objective: This paper aims to prompt large language models (LLMs) for clinical temporal relation extraction (CTRE) in both few-shot and fully supervised settings. Materials and Methods: This study utilizes four LLMs: Encoder-based GatorTron-Base (345M)/Large (8.9B); Decoder-based LLaMA3-8B/MeLLaMA-13B. We developed full (FFT) and parameter-efficient (PEFT) fine-tuning strategies and evaluated these strategies on the 2012 i2b2 CTRE task. We explored four fine-tuning strategies for GatorTron-Base: (1) Standard Fine-Tuning, (2) Hard-Prompting with Unfrozen LLMs, (3) Soft-Prompting with Frozen LLMs, and (4) Low-Rank Adaptation (LoRA) with Frozen LLMs. For GatorTron-Large, we assessed two PEFT strategies-Soft-Prompting and LoRA with Frozen LLMs-leveraging Quantization techniques. Additionally, LLaMA3-8B and MeLLaMA-13B employed two PEFT strategies: LoRA strategy with Quantization (QLoRA) applied to Frozen LLMs using instruction tuning and standard fine-tuning. Results: Under fully supervised settings, Hard-Prompting with Unfrozen GatorTron-Base achieved the highest F1 score (89.54%), surpassing the SOTA model (85.70%) by 3.74%. Additionally, two variants of QLoRA adapted to GatorTron-Large and Standard Fine-Tuning of GatorTron-Base exceeded the SOTA model by 2.36%, 1.88%, and 0.25%, respectively. Decoder-based models with frozen parameters outperformed their Encoder-based counterparts in this setting; however, the trend reversed in few-shot scenarios. Discussions and Conclusions: This study presented new methods that significantly improved CTRE performance, benefiting downstream tasks reliant on CTRE systems. The findings underscore the importance of selecting appropriate models and fine-tuning strategies based on task requirements and data availability. Future work will explore larger models and broader CTRE applications.
This study presents the conditional neural fields for reduced-order modeling (CNF-ROM) framework to approximate solutions of parametrized partial differential equations (PDEs). The approach combines a parametric neural ODE (PNODE) for modeling latent dynamics over time with a decoder that reconstructs PDE solutions from the corresponding latent states. We introduce a physics-informed learning objective for CNF-ROM, which includes two key components. First, the framework uses coordinate-based neural networks to calculate and minimize PDE residuals by computing spatial derivatives via automatic differentiation and applying the chain rule for time derivatives. Second, exact initial and boundary conditions (IC/BC) are imposed using approximate distance functions (ADFs) [Sukumar and Srivastava, CMAME, 2022]. However, ADFs introduce a trade-off as their second- or higher-order derivatives become unstable at the joining points of boundaries. To address this, we introduce an auxiliary network inspired by [Gladstone et al., NeurIPS ML4PS workshop, 2022]. Our method is validated through parameter extrapolation and interpolation, temporal extrapolation, and comparisons with analytical solutions.
We present a deep-learning Variational Encoder-Decoder (VED) framework for learning data-driven low-dimensional representations of the relationship between high-dimensional parameters of a physical system and the system's high-dimensional observable response. The framework consists of two deep learning-based probabilistic transformations: An encoder mapping parameters to latent codes and a decoder mapping latent codes to the observable response. The hyperparameters of these transformations are identified by maximizing a variational lower bound on the log-conditional distribution of the observable response given parameters. To promote the disentanglement of latent codes, we equip this variational loss with a penalty on the off-diagonal entries of the aggregate distribution covariance of codes. This regularization penalty encourages the pushforward of a standard Gaussian distribution of latent codes to approximate the marginal distribution of the observable response. Using the proposed framework we successfully model the hydraulic pressure response at observation wells of a groundwater flow model as a function of its discrete log-hydraulic transmissivity field. Compared to the canonical correlation analysis encoding, the VED model achieves a lower-dimensional latent representation, with as low as $r = 50$ latent dimensions without a significant loss of reconstruction accuracy. We explore the impact of regularization on model performance, finding that KL-divergence and covariance regularization improve feature disentanglement in latent space while maintaining reconstruction accuracy. Furthermore, we evaluate the generative capabilities of the regularized model by decoding random Gaussian noise, revealing that tuning both $\beta$ and $\lambda$ parameters enhances the quality of the generated observable response data.
We present UniMIC, a universal multi-modality image compression framework, intending to unify the rate-distortion-perception (RDP) optimization for multiple image codecs simultaneously through excavating cross-modality generative priors. Unlike most existing works that need to design and optimize image codecs from scratch, our UniMIC introduces the visual codec repository, which incorporates amounts of representative image codecs and directly uses them as the basic codecs for various practical applications. Moreover, we propose multi-grained textual coding, where variable-length content prompt and compression prompt are designed and encoded to assist the perceptual reconstruction through the multi-modality conditional generation. In particular, a universal perception compensator is proposed to improve the perception quality of decoded images from all basic codecs at the decoder side by reusing text-assisted diffusion priors from stable diffusion. With the cooperation of the above three strategies, our UniMIC achieves a significant improvement of RDP optimization for different compression codecs, e.g., traditional and learnable codecs, and different compression costs, e.g., ultra-low bitrates. The code will be available in https://github.com/Amygyx/UniMIC .
Recent advances in deep learning and natural language generation have significantly improved image captioning, enabling automated, human-like descriptions for visual content. In this work, we apply these captioning techniques to generate clinician-like interpretations of ECG data. This study leverages existing ECG datasets accompanied by free-text reports authored by healthcare professionals (HCPs) as training data. These reports, while often inconsistent, provide a valuable foundation for automated learning. We introduce an encoder-decoder-based method that uses these reports to train models to generate detailed descriptions of ECG episodes. This represents a significant advancement in ECG analysis automation, with potential applications in zero-shot classification and automated clinical decision support. The model is tested on various datasets, including both 1- and 12-lead ECGs. It significantly outperforms the state-of-the-art reference model by Qiu et al., achieving a METEOR score of 55.53% compared to 24.51% achieved by the reference model. Furthermore, several key design choices are discussed, providing a comprehensive overview of current challenges and innovations in this domain. The source codes for this research are publicly available in our Git repository https://git.zib.de/ableich/ecg-comment-generation-public
We introduce a novel method for human shape and pose recovery that can fully leverage multiple static views. We target fixed-multiview people monitoring, including elderly care and safety monitoring, in which calibrated cameras can be installed at the corners of a room or an open space but whose configuration may vary depending on the environment. Our key idea is to formulate it as neural optimization. We achieve this with HeatFormer, a neural optimizer that iteratively refines the SMPL parameters given multiview images, which is fundamentally agonistic to the configuration of views. HeatFormer realizes this SMPL parameter estimation as heat map generation and alignment with a novel transformer encoder and decoder. We demonstrate the effectiveness of HeatFormer including its accuracy, robustness to occlusion, and generalizability through an extensive set of experiments. We believe HeatFormer can serve a key role in passive human behavior modeling.
Despite the remarkable performance of multimodal large language models (MLLMs) across diverse tasks, the substantial training and inference costs impede their advancement. The majority of computation stems from the overwhelming volume of vision tokens processed by the transformer decoder. In this paper, we propose to build efficient MLLMs by leveraging the Mixture-of-Depths (MoD) mechanism, where each transformer decoder layer selects essential vision tokens to process while skipping redundant ones. However, integrating MoD into MLLMs is non-trivial. To address the challenges of training and inference stability as well as limited training data, we adapt the MoD module with two novel designs: tanh-gated weight normalization (TanhNorm) and symmetric token reweighting (STRing). Moreover, we observe that vision tokens exhibit higher redundancy in deeper layer and thus design a progressive ratio decay (PRD) strategy, which gradually reduces the token retention ratio layer by layer, employing a shifted cosine schedule. This crucial design fully unleashes the potential of MoD, significantly boosting the efficiency and performance of our models. To validate the effectiveness of our approach, we conduct extensive experiments with two baseline models across 14 benchmarks. Our model, p-MoD, matches or even surpasses the performance of the baseline models, with only 55.6% TFLOPs and 53.8% KV cache storage during inference, and 77.7% GPU hours during training.
Robotic-assisted minimally invasive esophagectomy (RAMIE) is a recognized treatment for esophageal cancer, offering better patient outcomes compared to open surgery and traditional minimally invasive surgery. RAMIE is highly complex, spanning multiple anatomical areas and involving repetitive phases and non-sequential phase transitions. Our goal is to leverage deep learning for surgical phase recognition in RAMIE to provide intraoperative support to surgeons. To achieve this, we have developed a new surgical phase recognition dataset comprising 27 videos. Using this dataset, we conducted a comparative analysis of state-of-the-art surgical phase recognition models. To more effectively capture the temporal dynamics of this complex procedure, we developed a novel deep learning model featuring an encoder-decoder structure with causal hierarchical attention, which demonstrates superior performance compared to existing models.
Transforming mathematical expressions into LaTeX poses a significant challenge. In this paper, we examine the application of advanced transformer-based architectures to address the task of converting handwritten or digital mathematical expression images into corresponding LaTeX code. As a baseline, we utilize the current state-of-the-art CNN encoder and LSTM decoder. Additionally, we explore enhancements to the CNN-RNN architecture by replacing the CNN encoder with the pretrained ResNet50 model with modification to suite the grey scale input. Further, we experiment with vision transformer model and compare with Baseline and CNN-LSTM model. Our findings reveal that the vision transformer architectures outperform the baseline CNN-RNN framework, delivering higher overall accuracy and BLEU scores while achieving lower Levenshtein distances. Moreover, these results highlight the potential for further improvement through fine-tuning of model parameters. To encourage open research, we also provide the model implementation, enabling reproduction of our results and facilitating further research in this domain.