The adoption of tablets with touchscreens and styluses is increasing, and a key feature is converting handwriting to text, enabling search, indexing, and AI assistance. Meanwhile, vision-language models (VLMs) are now the go-to solution for image understanding, thanks to both their state-of-the-art performance across a variety of tasks and the simplicity of a unified approach to training, fine-tuning, and inference. While VLMs obtain high performance on image-based tasks, they perform poorly on handwriting recognition when applied naively, i.e., by rendering handwriting as an image and performing optical character recognition (OCR). In this paper, we study online handwriting recognition with VLMs, going beyond naive OCR. We propose a novel tokenized representation of digital ink (online handwriting) that includes both a time-ordered sequence of strokes as text, and as image. We show that this representation yields results comparable to or better than state-of-the-art online handwriting recognizers. Wide applicability is shown through results with two different VLM families, on multiple public datasets. Our approach can be applied to off-the-shelf VLMs, does not require any changes in their architecture, and can be used in both fine-tuning and parameter-efficient tuning. We perform a detailed ablation study to identify the key elements of the proposed representation.
We propose a diffusion distillation method that achieves new state-of-the-art in one-step/few-step 1024px text-to-image generation based on SDXL. Our method combines progressive and adversarial distillation to achieve a balance between quality and mode coverage. In this paper, we discuss the theoretical analysis, discriminator design, model formulation, and training techniques. We open-source our distilled SDXL-Lightning models both as LoRA and full UNet weights.
In the physical sciences, there is an increased need for robust feature representations of image data: image acquisition, in the generalized sense of two-dimensional data, is now widespread across a large number of fields, including quantum information science, which we consider here. While traditional image features are widely utilized in such cases, their use is rapidly being supplanted by Neural Network-based techniques that often sacrifice explainability in exchange for high accuracy. To ameliorate this trade-off, we propose a synthetic data-based technique that results in explainable features. We show, using Explainable Boosting Machines (EBMs), that this method offers superior explainability without sacrificing accuracy. Specifically, we show that there is a meaningful benefit to this technique in the context of quantum dot tuning, where human intervention is necessary at the current stage of development.
Ultra-high resolution images are desirable in photon counting CT (PCCT), but resolution is physically limited by interactions such as charge sharing. Deep learning is a possible method for super-resolution (SR), but sourcing paired training data that adequately models the target task is difficult. Additionally, SR algorithms can distort noise texture, which is an important in many clinical diagnostic scenarios. Here, we train conditional denoising diffusion probabilistic models (DDPMs) for PCCT super-resolution, with the objective to retain textural characteristics of local noise. PCCT simulation methods are used to synthesize realistic resolution degradation. To preserve noise texture, we explore decoupling the noise and signal image inputs and outputs via deep denoisers, explicitly mapping to each during the SR process. Our experimental results indicate that our DDPM trained on simulated data can improve sharpness in real PCCT images. Additionally, the disentanglement of noise from the original image allows our model more faithfully preserve noise texture.
MoodCapture presents a novel approach that assesses depression based on images automatically captured from the front-facing camera of smartphones as people go about their daily lives. We collect over 125,000 photos in the wild from N=177 participants diagnosed with major depressive disorder for 90 days. Images are captured naturalistically while participants respond to the PHQ-8 depression survey question: \textit{``I have felt down, depressed, or hopeless''}. Our analysis explores important image attributes, such as angle, dominant colors, location, objects, and lighting. We show that a random forest trained with face landmarks can classify samples as depressed or non-depressed and predict raw PHQ-8 scores effectively. Our post-hoc analysis provides several insights through an ablation study, feature importance analysis, and bias assessment. Importantly, we evaluate user concerns about using MoodCapture to detect depression based on sharing photos, providing critical insights into privacy concerns that inform the future design of in-the-wild image-based mental health assessment tools.
Image coding for machines (ICM) aims to compress images for machine analysis using recognition models rather than human vision. Hence, in ICM, it is important for the encoder to recognize and compress the information necessary for the machine recognition task. There are two main approaches in learned ICM; optimization of the compression model based on task loss, and Region of Interest (ROI) based bit allocation. These approaches provide the encoder with the recognition capability. However, optimization with task loss becomes difficult when the recognition model is deep, and ROI-based methods often involve extra overhead during evaluation. In this study, we propose a novel training method for learned ICM models that applies auxiliary loss to the encoder to improve its recognition capability and rate-distortion performance. Our method achieves Bjontegaard Delta rate improvements of 27.7% and 20.3% in object detection and semantic segmentation tasks, compared to the conventional training method.
This paper presents ShapeLLM, the first 3D Multimodal Large Language Model (LLM) designed for embodied interaction, exploring a universal 3D object understanding with 3D point clouds and languages. ShapeLLM is built upon an improved 3D encoder by extending ReCon to ReCon++ that benefits from multi-view image distillation for enhanced geometry understanding. By utilizing ReCon++ as the 3D point cloud input encoder for LLMs, ShapeLLM is trained on constructed instruction-following data and tested on our newly human-curated evaluation benchmark, 3D MM-Vet. ReCon++ and ShapeLLM achieve state-of-the-art performance in 3D geometry understanding and language-unified 3D interaction tasks, such as embodied visual grounding.
As one of the most popular parameter-efficient fine-tuning (PEFT) methods, low-rank adaptation (LoRA) is commonly applied to fine-tune large language models (LLMs). However, updating the weights of LoRA blocks effectively and expeditiously is challenging due to the long calculation path in the original model. To address this, we propose ResLoRA, an improved framework of LoRA. By adding residual paths during training and using merging approaches to eliminate these extra paths during inference, our method can achieve better results in fewer training steps without any extra trainable parameters or inference cost compared to LoRA. The experiments on NLG, NLU, and text-to-image tasks demonstrate the effectiveness of our method. To the best of our knowledge, ResLoRA is the first work that combines the residual path with LoRA. The code of our method is available at https://github.com/microsoft/LMOps/tree/main/reslora .
In recent years, model quantization for face recognition has gained prominence. Traditionally, compressing models involved vast datasets like the 5.8 million-image MS1M dataset as well as extensive training times, raising the question of whether such data enormity is essential. This paper addresses this by introducing an efficiency-driven approach, fine-tuning the model with just up to 14,000 images, 440 times smaller than MS1M. We demonstrate that effective quantization is achievable with a smaller dataset, presenting a new paradigm. Moreover, we incorporate an evaluation-based metric loss and achieve an outstanding 96.15% accuracy on the IJB-C dataset, establishing a new state-of-the-art compressed model training for face recognition. The subsequent analysis delves into potential applications, emphasizing the transformative power of this approach. This paper advances model quantization by highlighting the efficiency and optimal results with small data and training time.
Robust quantification of pulmonary emphysema on computed tomography (CT) remains challenging for large-scale research studies that involve scans from different scanner types and for translation to clinical scans. Existing studies have explored several directions to tackle this challenge, including density correction, noise filtering, regression, hidden Markov measure field (HMMF) model-based segmentation, and volume-adjusted lung density. Despite some promising results, previous studies either required a tedious workflow or limited opportunities for downstream emphysema subtyping, limiting efficient adaptation on a large-scale study. To alleviate this dilemma, we developed an end-to-end deep learning framework based on an existing HMMF segmentation framework. We first demonstrate that a regular UNet cannot replicate the existing HMMF results because of the lack of scanner priors. We then design a novel domain attention block to fuse image feature with quantitative scanner priors which significantly improves the results.