The Segment Anything Model (SAM) is a powerful foundation model that introduced revolutionary advancements in natural image segmentation. However, its performance remains sub-optimal when delineating the intricate structure of biomedical images, where multiple organs and tissues intertwine in a single image. In this study, we introduce a novel fine-tuning framework that leverages SAM's ability to bundle and process multiple prompts per image and seeks to improve SAM's performance in medical images. We first curated a medical image dataset that consists of CT scans of lesions in various organs, each with two annotations for organs and lesions respectively. Then, we fine-tuned SAM's mask decoder within our framework by batching both bounding boxes generated from ground truth masks as reference. The batched prompt strategy we introduced not only addresses the inherent complexity and ambiguity often found in medical images but also substantially enhances performance metrics when applied onto a wide range of segmentation tasks.
Recent advancements in subject-driven image generation have made significant strides. However, current methods still fall short in diverse application scenarios, as they require test-time tuning and cannot accept interleaved multi-image and text input. These limitations keep them far from the ultimate goal of "image as a foreign language in image generation." This paper presents Kosmos-G, a model that leverages the advanced multimodal perception capabilities of Multimodal Large Language Models (MLLMs) to tackle the aforementioned challenge. Our approach aligns the output space of MLLM with CLIP using the textual modality as an anchor and performs compositional instruction tuning on curated data. Kosmos-G demonstrates an impressive capability of zero-shot subject-driven generation with interleaved multi-image and text input. Notably, the score distillation instruction tuning requires no modifications to the image decoder. This allows for a seamless substitution of CLIP and effortless integration with a myriad of U-Net techniques ranging from fine-grained controls to personalized image decoder variants. We posit Kosmos-G as an initial attempt towards the goal of "image as a foreign language in image generation." The code can be found at https://aka.ms/Kosmos-G
Transformer-based language models (LMs) are inefficient in long contexts. We propose Dodo, a solution for context compression. Instead of one vector per token in a standard transformer model, Dodo represents text with a dynamic number of hidden states at each layer, reducing the cost of self-attention to a fraction of typical time and space. Moreover, off-the-shelf models such as LLaMA can be adapted to Dodo by efficient parameter tuning methods such as LoRA. In use, Dodo can act as either an autoregressive LM or a context compressor for downstream tasks. We demonstrate through experiments in language modeling, question answering, and summarization that Dodo retains capabilities in these tasks, while drastically reducing the overhead during decoding. For example, in the autoencoding task, Dodo shrinks context at a 20x compression ratio with a BLEU score of 98% for reconstruction, achieving nearly lossless encoding.
Humans inherently possess generalizable visual representations that empower them to efficiently explore and interact with the environments in manipulation tasks. We advocate that such a representation automatically arises from simultaneously learning about multiple simple perceptual skills that are critical for everyday scenarios (e.g., hand detection, state estimate, etc.) and is better suited for learning robot manipulation policies compared to current state-of-the-art visual representations purely based on self-supervised objectives. We formalize this idea through the lens of human-oriented multi-task fine-tuning on top of pre-trained visual encoders, where each task is a perceptual skill tied to human-environment interactions. We introduce Task Fusion Decoder as a plug-and-play embedding translator that utilizes the underlying relationships among these perceptual skills to guide the representation learning towards encoding meaningful structure for what's important for all perceptual skills, ultimately empowering learning of downstream robotic manipulation tasks. Extensive experiments across a range of robotic tasks and embodiments, in both simulations and real-world environments, show that our Task Fusion Decoder consistently improves the representation of three state-of-the-art visual encoders including R3M, MVP, and EgoVLP, for downstream manipulation policy-learning. Project page: https://sites.google.com/view/human-oriented-robot-learning
Recently, neural networks have shown impressive progress across diverse fields, with speech processing being no exception. However, recent breakthroughs in this area require extensive offline training using large datasets and tremendous computing resources. Unfortunately, these models struggle to retain their previously acquired knowledge when learning new tasks continually, and retraining from scratch is almost always impractical. In this paper, we investigate the problem of learning sequence-to-sequence models for spoken language understanding in a class-incremental learning (CIL) setting and we propose COCONUT, a CIL method that relies on the combination of experience replay and contrastive learning. Through a modified version of the standard supervised contrastive loss applied only to the rehearsal samples, COCONUT preserves the learned representations by pulling closer samples from the same class and pushing away the others. Moreover, we leverage a multimodal contrastive loss that helps the model learn more discriminative representations of the new data by aligning audio and text features. We also investigate different contrastive designs to combine the strengths of the contrastive loss with teacher-student architectures used for distillation. Experiments on two established SLU datasets reveal the effectiveness of our proposed approach and significant improvements over the baselines. We also show that COCONUT can be combined with methods that operate on the decoder side of the model, resulting in further metrics improvements.
This paper presents a novel framework for joint speaker diarization (SD) and automatic speech recognition (ASR), named SLIDAR (sliding-window diarization-augmented recognition). SLIDAR can process arbitrary length inputs and can handle any number of speakers, effectively solving ``who spoke what, when'' concurrently. SLIDAR leverages a sliding window approach and consists of an end-to-end diarization-augmented speech transcription (E2E DAST) model which provides, locally, for each window: transcripts, diarization and speaker embeddings. The E2E DAST model is based on an encoder-decoder architecture and leverages recent techniques such as serialized output training and ``Whisper-style" prompting. The local outputs are then combined to get the final SD+ASR result by clustering the speaker embeddings to get global speaker identities. Experiments performed on monaural recordings from the AMI corpus confirm the effectiveness of the method in both close-talk and far-field speech scenarios.
This paper presents a classification framework based on learnable data augmentation to tackle the One-Shot Unsupervised Domain Adaptation (OS-UDA) problem. OS-UDA is the most challenging setting in Domain Adaptation, as only one single unlabeled target sample is assumed to be available for model adaptation. Driven by such single sample, our method LearnAug-UDA learns how to augment source data, making it perceptually similar to the target. As a result, a classifier trained on such augmented data will generalize well for the target domain. To achieve this, we designed an encoder-decoder architecture that exploits a perceptual loss and style transfer strategies to augment the source data. Our method achieves state-of-the-art performance on two well-known Domain Adaptation benchmarks, DomainNet and VisDA. The project code is available at https://github.com/IIT-PAVIS/LearnAug-UDA
Masked AutoEncoder (MAE) has revolutionized the field of self-supervised learning with its simple yet effective masking and reconstruction strategies. However, despite achieving state-of-the-art performance across various downstream vision tasks, the underlying mechanisms that drive MAE's efficacy are less well-explored compared to the canonical contrastive learning paradigm. In this paper, we first propose a local perspective to explicitly extract a local contrastive form from MAE's reconstructive objective at the patch level. And then we introduce a new empirical framework, called Local Contrastive MAE (LC-MAE), to analyze both reconstructive and contrastive aspects of MAE. LC-MAE reveals that MAE learns invariance to random masking and ensures distribution consistency between the learned token embeddings and the original images. Furthermore, we dissect the contribution of the decoder and random masking to MAE's success, revealing both the decoder's learning mechanism and the dual role of random masking as data augmentation and effective receptive field restriction. Our experimental analysis sheds light on the intricacies of MAE and summarizes some useful design methodologies, which can inspire more powerful visual self-supervised methods.
Current Large Multimodal Models (LMMs) achieve remarkable progress, yet there remains significant uncertainty regarding their ability to accurately apprehend visual details, that is, in performing detailed captioning. To address this, we introduce $\textit{CCEval}$, a GPT-4 assisted evaluation method for detailed captioning. Interestingly, while LMMs demonstrate minimal object existence hallucination in existing VQA benchmarks, our proposed evaluation reveals continued susceptibility to such hallucinations. In this paper, we make the first attempt to investigate such hallucination from different aspects, including image resolution, the language decoder size, and instruction data amount, quality, granularity. Our findings underscore the unwarranted inference when the language description includes details at a finer object granularity than what the vision module can ground or verify, thus inducing hallucination. To control such hallucinations, we further attribute the reliability of captioning to contextual knowledge (involving only contextually grounded objects) and parametric knowledge (containing inferred objects by the model). Thus, we introduce $\textit{HallE-Control}$, a controllable LMM in terms of $\textbf{Hall}$ucination in object $\textbf{E}$xistence. HallE-Control can condition the captioning to shift between (i) exclusively depicting contextual knowledge for grounded objects and (ii) blending it with parametric knowledge to imagine inferred objects. Our method reduces hallucination by 44% compared to LLaVA$_{7B}$ and maintains the object coverage.
Language models generate responses by producing a series of tokens in immediate succession: the $(K+1)^{th}$ token is an outcome of manipulating $K$ hidden vectors per layer, one vector per preceding token. What if instead we were to let the model manipulate say, $K+10$ hidden vectors, before it outputs the $(K+1)^{th}$ token? We operationalize this idea by performing training and inference on language models with a (learnable) $\textit{pause}$ token, a sequence of which is appended to the input prefix. We then delay extracting the model's outputs until the last pause token is seen, thereby allowing the model to process extra computation before committing to an answer. We empirically evaluate $\textit{pause-training}$ on decoder-only models of 1B and 130M parameters with causal pretraining on C4, and on downstream tasks covering reasoning, question-answering, general understanding and fact recall. Our main finding is that inference-time delays show gains when the model is both pre-trained and finetuned with delays. For the 1B model, we witness gains on 8 of 9 tasks, most prominently, a gain of $18\%$ EM score on the QA task of SQuAD, $8\%$ on CommonSenseQA and $1\%$ accuracy on the reasoning task of GSM8k. Our work raises a range of conceptual and practical future research questions on making delayed next-token prediction a widely applicable new paradigm.
De novo drug design requires simultaneously generating novel molecules outside of training data and predicting their target properties, making it a hard task for generative models. To address this, we propose Joint Transformer that combines a Transformer decoder, Transformer encoder, and a predictor in a joint generative model with shared weights. We formulate a probabilistic black-box optimization algorithm that employs Joint Transformer to generate novel molecules with improved target properties and outperforms other SMILES-based optimization methods in de novo drug design.
We present PARQ - a multi-view 3D object detector with transformer and pixel-aligned recurrent queries. Unlike previous works that use learnable features or only encode 3D point positions as queries in the decoder, PARQ leverages appearance-enhanced queries initialized from reference points in 3D space and updates their 3D location with recurrent cross-attention operations. Incorporating pixel-aligned features and cross attention enables the model to encode the necessary 3D-to-2D correspondences and capture global contextual information of the input images. PARQ outperforms prior best methods on the ScanNet and ARKitScenes datasets, learns and detects faster, is more robust to distribution shifts in reference points, can leverage additional input views without retraining, and can adapt inference compute by changing the number of recurrent iterations.
Neural video codecs have recently become competitive with standard codecs such as HEVC in the low-delay setting. However, most neural codecs are large floating-point networks that use pixel-dense warping operations for temporal modeling, making them too computationally expensive for deployment on mobile devices. Recent work has demonstrated that running a neural decoder in real time on mobile is feasible, but shows this only for 720p RGB video. This work presents the first neural video codec that decodes 1080p YUV420 video in real time on a mobile device. Our codec relies on two major contributions. First, we design an efficient codec that uses a block-based motion compensation algorithm available on the warping core of the mobile accelerator, and we show how to quantize this model to integer precision. Second, we implement a fast decoder pipeline that concurrently runs neural network components on the neural signal processor, parallel entropy coding on the mobile GPU, and warping on the warping core. Our codec outperforms the previous on-device codec by a large margin with up to 48% BD-rate savings, while reducing the MAC count on the receiver side by $10 \times$. We perform a careful ablation to demonstrate the effect of the introduced motion compensation scheme, and ablate the effect of model quantization.
We propose an end-to-end learned image data hiding framework that embeds and extracts secrets in the latent representations of a generic neural compressor. By leveraging a perceptual loss function in conjunction with our proposed message encoder and decoder, our approach simultaneously achieves high image quality and high bit accuracy. Compared to existing techniques, our framework offers superior image secrecy and competitive watermarking robustness in the compressed domain while accelerating the embedding speed by over 50 times. These results demonstrate the potential of combining data hiding techniques and neural compression and offer new insights into developing neural compression techniques and their applications.
Structured documents analysis and recognition are essential for modern online on-boarding processes, and document localization is a crucial step to achieve reliable key information extraction. While deep-learning has become the standard technique used to solve document analysis problems, real-world applications in industry still face the limited availability of labelled data and of computational resources when training or fine-tuning deep-learning models. To tackle these challenges, we propose SDL-Net: a novel U-Net like encoder-decoder architecture for the localization of structured documents. Our approach allows pre-training the encoder of SDL-Net on a generic dataset containing samples of various document classes, and enables fast and data-efficient fine-tuning of decoders to support the localization of new document classes. We conduct extensive experiments on a proprietary dataset of structured document images to demonstrate the effectiveness and the generalization capabilities of the proposed approach.
For decades, fingerprint recognition has been prevalent for security, forensics, and other biometric applications. However, the availability of good-quality fingerprints is challenging, making recognition difficult. Fingerprint images might be degraded with a poor ridge structure and noisy or less contrasting backgrounds. Hence, fingerprint enhancement plays a vital role in the early stages of the fingerprint recognition/verification pipeline. In this paper, we investigate and improvise the encoder-decoder style architecture and suggest intuitive modifications to U-Net to enhance low-quality fingerprints effectively. We investigate the use of Discrete Wavelet Transform (DWT) for fingerprint enhancement and use a wavelet attention module instead of max pooling which proves advantageous for our task. Moreover, we replace regular convolutions with depthwise separable convolutions, which significantly reduces the memory footprint of the model without degrading the performance. We also demonstrate that incorporating domain knowledge with fingerprint minutiae prediction task can improve fingerprint reconstruction through multi-task learning. Furthermore, we also integrate the orientation estimation task to propagate the knowledge of ridge orientations to enhance the performance further. We present the experimental results and evaluate our model on FVC 2002 and NIST SD302 databases to show the effectiveness of our approach compared to previous works.
We proposed an easy method of Zero-Shot semantic segmentation by using style transfer. In this case, we successfully used a medical imaging dataset (Blood Cell Imagery) to train a model for river ice semantic segmentation. First, we built a river ice semantic segmentation dataset IPC_RI_SEG using a fixed camera and covering the entire ice melting process of the river. Second, a high-resolution texture fusion semantic segmentation network named IceHrNet is proposed. The network used HRNet as the backbone and added ASPP and Decoder segmentation heads to retain low-level texture features for fine semantic segmentation. Finally, a simple and effective advanced style transfer learning strategy was proposed, which can perform zero-shot transfer learning based on cross-domain semantic segmentation datasets, achieving a practical effect of 87% mIoU for semantic segmentation of river ice without target training dataset (25% mIoU for None Stylized, 65% mIoU for Conventional Stylized, our strategy improved by 22%). Experiments showed that the IceHrNet outperformed the state-of-the-art methods on the texture-focused dataset IPC_RI_SEG, and achieved an excellent result on the shape-focused river ice datasets. In zero-shot transfer learning, IceHrNet achieved an increase of 2 percentage points compared to other methods. Our code and model are published on https://github.com/PL23K/IceHrNet.
We propose an automatic segmentation method for lumen and media with irregular contours in IntraVascular ultra-sound (IVUS) images. In contrast to most approaches that broadly label each pixel as either lumen, media, or background, we propose to approximate the lumen and media contours by closed polygonal chains. The chain vertices are placed at fixed angles obtained by dividing the entire 360\degree~angular space into equally spaced angles, and we predict their radius using an adaptive-subband-decomposition CNN. We consider two loss functions during training. The first is a novel loss function using the Jaccard Measure (JM) to quantify the similarities between the predicted lumen and media segments and the corresponding ground-truth image segments. The second loss function is the traditional Mean Squared Error. The proposed architecture significantly reduces computational costs by replacing the popular auto-encoder structure with a simple CNN as the encoder and the decoder is reduced to simply joining the consecutive predicted points. We evaluated our network on the publicly available IVUS-Challenge-2011 dataset using two performance metrics, namely JM and Hausdorff Distance (HD). The evaluation results show that our proposed network mostly outperforms the state-of-the-art lumen and media segmentation methods.
Quantitative cardiac magnetic resonance T1 and T2 mapping enable myocardial tissue characterisation but the lengthy scan times restrict their widespread clinical application. We propose a deep learning method that incorporates a time dependency Latent Transformer module to model relationships between parameterised time frames for improved reconstruction from undersampled data. The module, implemented as a multi-resolution sequence-to-sequence transformer, is integrated into an encoder-decoder architecture to leverage the inherent temporal correlations in relaxation processes. The presented results for accelerated T1 and T2 mapping show the model recovers maps with higher fidelity by explicit incorporation of time dynamics. This work demonstrates the importance of temporal modelling for artifact-free reconstruction in quantitative MRI.
Hallucination in text summarization refers to the phenomenon where the model generates information that is not supported by the input source document. Hallucination poses significant obstacles to the accuracy and reliability of the generated summaries. In this paper, we aim to reduce hallucinated outputs or hallucinations in summaries of long-form text documents. We have used the PubMed dataset, which contains long scientific research documents and their abstracts. We have incorporated the techniques of data filtering and joint entity and summary generation (JAENS) in the fine-tuning of the Longformer Encoder-Decoder (LED) model to minimize hallucinations and thereby improve the quality of the generated summary. We have used the following metrics to measure factual consistency at the entity level: precision-source, and F1-target. Our experiments show that the fine-tuned LED model performs well in generating the paper abstract. Data filtering techniques based on some preprocessing steps reduce entity-level hallucinations in the generated summaries in terms of some of the factual consistency metrics.