We introduce MobileVLM V2, a family of significantly improved vision language models upon MobileVLM, which proves that a delicate orchestration of novel architectural design, an improved training scheme tailored for mobile VLMs, and rich high-quality dataset curation can substantially benefit VLMs' performance. Specifically, MobileVLM V2 1.7B achieves better or on-par performance on standard VLM benchmarks compared with much larger VLMs at the 3B scale. Notably, our 3B model outperforms a large variety of VLMs at the 7B+ scale. Our models will be released at https://github.com/Meituan-AutoML/MobileVLM .
We present MobileVLM, a competent multimodal vision language model (MMVLM) targeted to run on mobile devices. It is an amalgamation of a myriad of architectural designs and techniques that are mobile-oriented, which comprises a set of language models at the scale of 1.4B and 2.7B parameters, trained from scratch, a multimodal vision model that is pre-trained in the CLIP fashion, cross-modality interaction via an efficient projector. We evaluate MobileVLM on several typical VLM benchmarks. Our models demonstrate on par performance compared with a few much larger models. More importantly, we measure the inference speed on both a Qualcomm Snapdragon 888 CPU and an NVIDIA Jeston Orin GPU, and we obtain state-of-the-art performance of 21.5 tokens and 65.3 tokens per second, respectively. Our code will be made available at: https://github.com/Meituan-AutoML/MobileVLM.
Image fusion aims to combine information from multiple source images into a single and more informative image. A major challenge for deep learning-based image fusion algorithms is the absence of a definitive ground truth and distance measurement. Thus, the manually specified loss functions aiming to steer the model learning, include hyperparameters that need to be manually thereby limiting the model's flexibility and generalizability to unseen tasks. To overcome the limitations of designing loss functions for specific fusion tasks, we propose a unified meta-learning based fusion framework named ReFusion, which learns optimal fusion loss from reconstructing source images. ReFusion consists of a fusion module, a loss proposal module, and a reconstruction module. Compared with the conventional methods with fixed loss functions, ReFusion employs a parameterized loss function, which is dynamically adapted by the loss proposal module based on the specific fusion scene and task. To ensure that the fusion network preserves maximal information from the source images, makes it possible to reconstruct the original images from the fusion image, a meta-learning strategy is used to make the reconstruction loss continually refine the parameters of the loss proposal module. Adaptive updating is achieved by alternating between inter update, outer update, and fusion update, where the training of the three components facilitates each other. Extensive experiments affirm that our method can successfully adapt to diverse fusion tasks, including infrared-visible, multi-focus, multi-exposure, and medical image fusion problems. The code will be released.
Current traditional methods for LiDAR-camera extrinsics estimation depend on offline targets and human efforts, while learning-based approaches resort to iterative refinement for calibration results, posing constraints on their generalization and application in on-board systems. In this paper, we propose a novel approach to address the extrinsic calibration problem in a robust, automatic, and single-shot manner. Instead of directly optimizing extrinsics, we leverage the consistency learning between LiDAR and camera to implement implicit re-calibartion. Specially, we introduce an appearance-consistency loss and a geometric-consistency loss to minimizing the inconsitency between the attrbutes (e.g., intensity and depth) of projected LiDAR points and the predicted ones. This design not only enhances adaptability to various scenarios but also enables a simple and efficient formulation during inference. We conduct comprehensive experiments on different datasets, and the results demonstrate that our method achieves accurate and robust performance. To promote further research and development in this area, we will release our model and code.
Owing to its significant success, the prior imposed on gradient maps has consistently been a subject of great interest in the field of image processing. Total variation (TV), one of the most representative regularizers, is known for its ability to capture the intrinsic sparsity prior underlying gradient maps. Nonetheless, TV and its variants often underestimate the gradient maps, leading to the weakening of edges and details whose gradients should not be zero in the original image (i.e., image structures is not describable by sparse priors of gradient maps). Recently, total deep variation (TDV) has been introduced, assuming the sparsity of feature maps, which provides a flexible regularization learned from large-scale datasets for a specific task. However, TDV requires to retrain the network with image/task variations, limiting its versatility. To alleviate this issue, in this paper, we propose a neural gradient regularizer (NGR) that expresses the gradient map as the output of a neural network. Unlike existing methods, NGR does not rely on any subjective sparsity or other prior assumptions on image gradient maps, thereby avoiding the underestimation of gradient maps. NGR is applicable to various image types and different image processing tasks, functioning in a zero-shot learning fashion, making it a versatile and plug-and-play regularizer. Extensive experimental results demonstrate the superior performance of NGR over state-of-the-art counterparts for a range of different tasks, further validating its effectiveness and versatility.
This paper focuses on term-status pair extraction from medical dialogues (MD-TSPE), which is essential in diagnosis dialogue systems and the automatic scribe of electronic medical records (EMRs). In the past few years, works on MD-TSPE have attracted increasing research attention, especially after the remarkable progress made by generative methods. However, these generative methods output a whole sequence consisting of term-status pairs in one stage and ignore integrating prior knowledge, which demands a deeper understanding to model the relationship between terms and infer the status of each term. This paper presents a knowledge-enhanced two-stage generative framework (KTGF) to address the above challenges. Using task-specific prompts, we employ a single model to complete the MD-TSPE through two phases in a unified generative form: we generate all terms the first and then generate the status of each generated term. In this way, the relationship between terms can be learned more effectively from the sequence containing only terms in the first phase, and our designed knowledge-enhanced prompt in the second phase can leverage the category and status candidates of the generated term for status generation. Furthermore, our proposed special status ``not mentioned" makes more terms available and enriches the training data in the second phase, which is critical in the low-resource setting. The experiments on the Chunyu and CMDD datasets show that the proposed method achieves superior results compared to the state-of-the-art models in the full training and low-resource settings.
Hyperspectral image (HSI) denoising is essentially ill-posed since a noisy HSI can be degraded from multiple clean HSIs. However, current deep learning-based approaches ignore this fact and restore the clean image with deterministic mapping (i.e., the network receives a noisy HSI and outputs a clean HSI). To alleviate this issue, this paper proposes a flow-based HSI denoising network (HIDFlowNet) to directly learn the conditional distribution of the clean HSI given the noisy HSI and thus diverse clean HSIs can be sampled from the conditional distribution. Overall, our HIDFlowNet is induced from the flow methodology and contains an invertible decoder and a conditional encoder, which can fully decouple the learning of low-frequency and high-frequency information of HSI. Specifically, the invertible decoder is built by staking a succession of invertible conditional blocks (ICBs) to capture the local high-frequency details since the invertible network is information-lossless. The conditional encoder utilizes down-sampling operations to obtain low-resolution images and uses transformers to capture correlations over a long distance so that global low-frequency information can be effectively extracted. Extensive experimental results on simulated and real HSI datasets verify the superiority of our proposed HIDFlowNet compared with other state-of-the-art methods both quantitatively and visually.
Employing additional multimodal information to improve automatic speech recognition (ASR) performance has been proven effective in previous works. However, many of these works focus only on the utilization of visual cues from human lip motion. In fact, context-dependent visual and linguistic cues can also be used to improve ASR performance in many scenarios. In this paper, we first propose a multimodal ASR model (ViLaS) that can simultaneously or separately integrate visual and linguistic cues to help recognize the input speech, and introduce a training strategy that can improve performance in modal-incomplete test scenarios. Then, we create a multimodal ASR dataset (VSDial) with visual and linguistic cues to explore the effects of integrating vision and language. Finally, we report empirical results on the public Flickr8K and self-constructed VSDial datasets, investigate cross-modal fusion schemes, and analyze fine-grained cross-modal alignment on VSDial.
Multi-modality image fusion is a technique used to combine information from different sensors or modalities, allowing the fused image to retain complementary features from each modality, such as functional highlights and texture details. However, effectively training such fusion models is difficult due to the lack of ground truth fusion data. To address this issue, we propose the Equivariant Multi-Modality imAge fusion (EMMA) paradigm for end-to-end self-supervised learning. Our approach is based on the prior knowledge that natural images are equivariant to specific transformations. Thus, we introduce a novel training framework that includes a fusion module and a learnable pseudo-sensing module, which allow the network training to follow the principles of physical sensing and imaging process, and meanwhile satisfy the equivariant prior for natural images. Our extensive experiments demonstrate that our method produces high-quality fusion results for both infrared-visible and medical images, while facilitating downstream multi-modal segmentation and detection tasks. The code will be released.