CompCap: Improving Multimodal Large Language Models with Composite Captions

How well can Multimodal Large Language Models (MLLMs) understand composite images? Composite images (CIs) are synthetic visuals created by merging multiple visual elements, such as charts, posters, or screenshots, rather than being captured directly by a camera. While CIs are prevalent in real-world applications, recent MLLM developments have primarily focused on interpreting natural images (NIs). Our research reveals that current MLLMs face significant challenges in accurately understanding CIs, often struggling to extract information or perform complex reasoning based on these images. We find that existing training data for CIs are mostly formatted for question-answer tasks (e.g., in datasets like ChartQA and ScienceQA), while high-quality image-caption datasets, critical for robust vision-language alignment, are only available for NIs. To bridge this gap, we introduce Composite Captions (CompCap), a flexible framework that leverages Large Language Models (LLMs) and automation tools to synthesize CIs with accurate and detailed captions. Using CompCap, we curate CompCap-118K, a dataset containing 118K image-caption pairs across six CI types. We validate the effectiveness of CompCap-118K by supervised fine-tuning MLLMs of three sizes: xGen-MM-inst.-4B and LLaVA-NeXT-Vicuna-7B/13B. Empirical results show that CompCap-118K significantly enhances MLLMs' understanding of CIs, yielding average gains of 1.7%, 2.0%, and 2.9% across eleven benchmarks, respectively.

Practical Phase Retrieval Using Double Deep Image Priors

Phase retrieval (PR) concerns the recovery of complex phases from complex magnitudes. We identify the connection between the difficulty level and the number and variety of symmetries in PR problems. We focus on the most difficult far-field PR (FFPR), and propose a novel method using double deep image priors. In realistic evaluation, our method outperforms all competing methods by large margins. As a single-instance method, our method requires no training data and minimal hyperparameter tuning, and hence enjoys good practicality.

Object-Centric Unsupervised Image Captioning

Training an image captioning model in an unsupervised manner without utilizing annotated image-caption pairs is an important step towards tapping into a wider corpus of text and images. In the supervised setting, image-caption pairs are "well-matched", where all objects mentioned in the sentence appear in the corresponding image. These pairings are, however, not available in the unsupervised setting. To overcome this, a main school of research that has been shown to be effective in overcoming this is to construct pairs from the images and texts in the training set according to their overlap of objects. Unlike in the supervised setting, these constructed pairings are however not guaranteed to have fully overlapping set of objects. Our work in this paper overcomes this by harvesting objects corresponding to a given sentence from the training set, even if they don't belong to the same image. When used as input to a transformer, such mixture of objects enable larger if not full object coverage, and when supervised by the corresponding sentence, produced results that outperform current state of the art unsupervised methods by a significant margin. Building upon this finding, we further show that (1) additional information on relationship between objects and attributes of objects also helps in boosting performance; and (2) our method also extends well to non-English image captioning, which usually suffers from a scarcer level of annotations. Our findings are supported by strong empirical results.

Phase Retrieval using Single-Instance Deep Generative Prior

Several deep learning methods for phase retrieval exist, but most of them fail on realistic data without precise support information. We propose a novel method based on single-instance deep generative prior that works well on complex-valued crystal data.

Multimodal Fusion Refiner Networks

Tasks that rely on multi-modal information typically include a fusion module that combines information from different modalities. In this work, we develop a Refiner Fusion Network (ReFNet) that enables fusion modules to combine strong unimodal representation with strong multimodal representations. ReFNet combines the fusion network with a decoding/defusing module, which imposes a modality-centric responsibility condition. This approach addresses a big gap in existing multimodal fusion frameworks by ensuring that both unimodal and fused representations are strongly encoded in the latent fusion space. We demonstrate that the Refiner Fusion Network can improve upon performance of powerful baseline fusion modules such as multimodal transformers. The refiner network enables inducing graphical representations of the fused embeddings in the latent space, which we prove under certain conditions and is supported by strong empirical results in the numerical experiments. These graph structures are further strengthened by combining the ReFNet with a Multi-Similarity contrastive loss function. The modular nature of Refiner Fusion Network lends itself to be combined with different fusion architectures easily, and in addition, the refiner step can be applied for pre-training on unlabeled datasets, thus leveraging unsupervised data towards improving performance. We demonstrate the power of Refiner Fusion Networks on three datasets, and further show that they can maintain performance with only a small fraction of labeled data.

Robust Deep Learning with Active Noise Cancellation for Spatial Computing

This paper proposes CANC, a Co-teaching Active Noise Cancellation method, applied in spatial computing to address deep learning trained with extreme noisy labels. Deep learning algorithms have been successful in spatial computing for land or building footprint recognition. However a lot of noise exists in ground truth labels due to how labels are collected in spatial computing and satellite imagery. Existing methods to deal with extreme label noise conduct clean sample selection and do not utilize the remaining samples. Such techniques can be wasteful due to the cost of data retrieval. Our proposed CANC algorithm not only conserves high-cost training samples but also provides active label correction to better improve robust deep learning with extreme noisy labels. We demonstrate the effectiveness of CANC for building footprint recognition for spatial computing.