Egocentric Interactive hand-object segmentation (EgoIHOS) requires the segmentation of hands and interacting objects in egocentric images, which is crucial for understanding human behavior in assistive systems. Previous methods typically recognize hands and interacting objects as distinct semantic categories based solely on visual features, or simply use hand predictions as auxiliary cues for object segmentation. Despite the promising progress achieved by these methods, they fail to adequately model the interactive relationships between hands and objects while ignoring the coupled physical relationships among object categories, ultimately constraining their segmentation performance. To make up for the shortcomings of existing methods, we propose a novel method called CaRe-Ego that achieves state-of-the-art performance by emphasizing the contact between hands and objects from two aspects. First, we introduce a Hand-guided Object Feature Enhancer (HOFE) to establish the hand-object interactive relationships to extract more contact-relevant and discriminative object features. Second, we design the Contact-centric Object Decoupling Strategy (CODS) to explicitly model and disentangle coupling relationships among object categories, thereby emphasizing contact-aware feature learning. Experiments on various in-domain and out-of-domain test sets show that Care-Ego significantly outperforms existing methods with robust generalization capability. Codes are publicly available at https://github.com/yuggiehk/CaRe-Ego/.
Inferring future activity information based on observed activity data is a crucial step to improve the accuracy of early activity prediction. Traditional methods based on generative adversarial networks(GAN) or joint learning frameworks can achieve good prediction accuracy under low observation ratios, but they usually have high computational costs. In view of this, this paper proposes a spatio-temporal encoding and decoding-based method for future human activity skeleton synthesis. Firstly, algorithms such as time control, discrete cosine transform, and low-pass filtering are used to cut or pad the skeleton sequences. Secondly, the encoder and decoder are responsible for extracting intermediate semantic encoding from observed skeleton sequences and inferring future sequences from the intermediate semantic encoding, respectively. Finally, joint displacement error, velocity error, and acceleration error, three higher-order kinematic features, are used as key components of the loss function to optimize model parameters. Experimental results show that the proposed future skeleton synthesis algorithm performs better than some existing algorithms. It generates skeleton sequences with smaller errors and fewer model parameters, effectively providing future information for early activity prediction.
Semi-supervised learning (SSL) has achieved notable progress in medical image segmentation. To achieve effective SSL, a model needs to be able to efficiently learn from limited labeled data and effectively exploiting knowledge from abundant unlabeled data. Recent developments in visual foundation models, such as the Segment Anything Model (SAM), have demonstrated remarkable adaptability with improved sample efficiency. To harness the power of foundation models for application in SSL, we propose a cross prompting consistency method with segment anything model (CPC-SAM) for semi-supervised medical image segmentation. Our method employs SAM's unique prompt design and innovates a cross-prompting strategy within a dual-branch framework to automatically generate prompts and supervisions across two decoder branches, enabling effectively learning from both scarce labeled and valuable unlabeled data. We further design a novel prompt consistency regularization, to reduce the prompt position sensitivity and to enhance the output invariance under different prompts. We validate our method on two medical image segmentation tasks. The extensive experiments with different labeled-data ratios and modalities demonstrate the superiority of our proposed method over the state-of-the-art SSL methods, with more than 9% Dice improvement on the breast cancer segmentation task.
World models are receiving increasing attention in autonomous driving for their ability to predict potential future scenarios. In this paper, we present BEVWorld, a novel approach that tokenizes multimodal sensor inputs into a unified and compact Bird's Eye View (BEV) latent space for environment modeling. The world model consists of two parts: the multi-modal tokenizer and the latent BEV sequence diffusion model. The multi-modal tokenizer first encodes multi-modality information and the decoder is able to reconstruct the latent BEV tokens into LiDAR and image observations by ray-casting rendering in a self-supervised manner. Then the latent BEV sequence diffusion model predicts future scenarios given action tokens as conditions. Experiments demonstrate the effectiveness of BEVWorld in autonomous driving tasks, showcasing its capability in generating future scenes and benefiting downstream tasks such as perception and motion prediction. Code will be available at https://github.com/zympsyche/BevWorld.
The development of domain-specific language models has significantly advanced natural language processing applications in various specialized fields, particularly in biomedicine. However, the focus has largely been on English-language models, leaving a gap for less-resourced languages such as Italian. This paper introduces Igea, the first decoder-only language model designed explicitly for biomedical text generation in Italian. Built on the Minerva model and continually pretrained on a diverse corpus of Italian medical texts, Igea is available in three model sizes: 350 million, 1 billion, and 3 billion parameters. The models aim to balance computational efficiency and performance, addressing the challenges of managing the peculiarities of medical terminology in Italian. We evaluate Igea using a mix of in-domain biomedical corpora and general-purpose benchmarks, highlighting its efficacy and retention of general knowledge even after the domain-specific training. This paper discusses the model's development and evaluation, providing a foundation for future advancements in Italian biomedical NLP.
Continual Learning (CL) involves fine-tuning pre-trained models with new data while maintaining the performance on the pre-trained data. This is particularly relevant for expanding multilingual ASR (MASR) capabilities. However, existing CL methods, mainly designed for computer vision and reinforcement learning tasks, often yield sub-optimal results when directly applied to MASR. We hypothesise that this is because CL of the auto-regressive decoder in the MASR model is difficult. To verify this, we propose four optimizations on the decoder. They include decoder-layer gradient surgery, freezing unused token embeddings, suppressing output of newly added tokens, and learning rate re-scaling. Our experiments on adapting Whisper to 10 unseen languages from the Common Voice dataset demonstrate that these optimizations reduce the Average Word Error Rate (AWER) of pretrained languages from 14.2% to 12.4% compared with Experience Replay, without compromising the AWER of new languages.
Two competing paradigms exist for self-supervised learning of data representations. Joint Embedding Predictive Architecture (JEPA) is a class of architectures in which semantically similar inputs are encoded into representations that are predictive of each other. A recent successful approach that falls under the JEPA framework is self-distillation, where an online encoder is trained to predict the output of the target encoder, sometimes using a lightweight predictor network. This is contrasted with the Masked AutoEncoder (MAE) paradigm, where an encoder and decoder are trained to reconstruct missing parts of the input in the data space rather, than its latent representation. A common motivation for using the JEPA approach over MAE is that the JEPA objective prioritizes abstract features over fine-grained pixel information (which can be unpredictable and uninformative). In this work, we seek to understand the mechanism behind this empirical observation by analyzing the training dynamics of deep linear models. We uncover a surprising mechanism: in a simplified linear setting where both approaches learn similar representations, JEPAs are biased to learn high-influence features, i.e., features characterized by having high regression coefficients. Our results point to a distinct implicit bias of predicting in latent space that may shed light on its success in practice.
Graph pre-training has been concentrated on graph-level on small graphs (e.g., molecular graphs) or learning node representations on a fixed graph. Extending graph pre-trained models to web-scale graphs with billions of nodes in industrial scenarios, while avoiding negative transfer across graphs or tasks, remains a challenge. We aim to develop a general graph pre-trained model with inductive ability that can make predictions for unseen new nodes and even new graphs. In this work, we introduce a scalable transformer-based graph pre-training framework called PGT (Pre-trained Graph Transformer). Specifically, we design a flexible and scalable graph transformer as the backbone network. Meanwhile, based on the masked autoencoder architecture, we design two pre-training tasks: one for reconstructing node features and the other one for reconstructing local structures. Unlike the original autoencoder architecture where the pre-trained decoder is discarded, we propose a novel strategy that utilizes the decoder for feature augmentation. We have deployed our framework on Tencent's online game data. Extensive experiments have demonstrated that our framework can perform pre-training on real-world web-scale graphs with over 540 million nodes and 12 billion edges and generalizes effectively to unseen new graphs with different downstream tasks. We further conduct experiments on the publicly available ogbn-papers100M dataset, which consists of 111 million nodes and 1.6 billion edges. Our framework achieves state-of-the-art performance on both industrial datasets and public datasets, while also enjoying scalability and efficiency.
Rendering high-fidelity images from sparse point clouds is still challenging. Existing learning-based approaches suffer from either hole artifacts, missing details, or expensive computations. In this paper, we propose a novel framework to render high-quality images from sparse points. This method first attempts to bridge the 3D Gaussian Splatting and point cloud rendering, which includes several cascaded modules. We first use a regressor to estimate Gaussian properties in a point-wise manner, the estimated properties are used to rasterize neural feature descriptors into 2D planes which are extracted from a multiscale extractor. The projected feature volume is gradually decoded toward the final prediction via a multiscale and progressive decoder. The whole pipeline experiences a two-stage training and is driven by our well-designed progressive and multiscale reconstruction loss. Experiments on different benchmarks show the superiority of our method in terms of rendering qualities and the necessities of our main components.
Image restoration is a classic low-level problem aimed at recovering high-quality images from low-quality images with various degradations such as blur, noise, rain, haze, etc. However, due to the inherent complexity and non-uniqueness of degradation in real-world images, it is challenging for a model trained for single tasks to handle real-world restoration problems effectively. Moreover, existing methods often suffer from over-smoothing and lack of realism in the restored results. To address these issues, we propose Diff-Restorer, a universal image restoration method based on the diffusion model, aiming to leverage the prior knowledge of Stable Diffusion to remove degradation while generating high perceptual quality restoration results. Specifically, we utilize the pre-trained visual language model to extract visual prompts from degraded images, including semantic and degradation embeddings. The semantic embeddings serve as content prompts to guide the diffusion model for generation. In contrast, the degradation embeddings modulate the Image-guided Control Module to generate spatial priors for controlling the spatial structure of the diffusion process, ensuring faithfulness to the original image. Additionally, we design a Degradation-aware Decoder to perform structural correction and convert the latent code to the pixel domain. We conducted comprehensive qualitative and quantitative analysis on restoration tasks with different degradations, demonstrating the effectiveness and superiority of our approach.
We present a novel framework for reconstructing animatable human avatars from multiple images, termed CanonicalFusion. Our central concept involves integrating individual reconstruction results into the canonical space. To be specific, we first predict Linear Blend Skinning (LBS) weight maps and depth maps using a shared-encoder-dual-decoder network, enabling direct canonicalization of the 3D mesh from the predicted depth maps. Here, instead of predicting high-dimensional skinning weights, we infer compressed skinning weights, i.e., 3-dimensional vector, with the aid of pre-trained MLP networks. We also introduce a forward skinning-based differentiable rendering scheme to merge the reconstructed results from multiple images. This scheme refines the initial mesh by reposing the canonical mesh via the forward skinning and by minimizing photometric and geometric errors between the rendered and the predicted results. Our optimization scheme considers the position and color of vertices as well as the joint angles for each image, thereby mitigating the negative effects of pose errors. We conduct extensive experiments to demonstrate the effectiveness of our method and compare our CanonicalFusion with state-of-the-art methods. Our source codes are available at https://github.com/jsshin98/CanonicalFusion.
This report presents a novel Mamba-based encoder-decoder architecture, QueryMamba, featuring an integrated verb-noun interaction module that utilizes a statistical verb-noun co-occurrence matrix to enhance video action forecasting. This architecture not only predicts verbs and nouns likely to occur based on historical data but also considers their joint occurrence to improve forecast accuracy. The efficacy of this approach is substantiated by experimental results, with the method achieving second place in the Ego4D LTA challenge and ranking first in noun prediction accuracy.
We propose a straightforward yet highly effective few-shot fine-tuning strategy for adapting the Segment Anything (SAM) to anatomical segmentation tasks in medical images. Our novel approach revolves around reformulating the mask decoder within SAM, leveraging few-shot embeddings derived from a limited set of labeled images (few-shot collection) as prompts for querying anatomical objects captured in image embeddings. This innovative reformulation greatly reduces the need for time-consuming online user interactions for labeling volumetric images, such as exhaustively marking points and bounding boxes to provide prompts slice by slice. With our method, users can manually segment a few 2D slices offline, and the embeddings of these annotated image regions serve as effective prompts for online segmentation tasks. Our method prioritizes the efficiency of the fine-tuning process by exclusively training the mask decoder through caching mechanisms while keeping the image encoder frozen. Importantly, this approach is not limited to volumetric medical images, but can generically be applied to any 2D/3D segmentation task. To thoroughly evaluate our method, we conducted extensive validation on four datasets, covering six anatomical segmentation tasks across two modalities. Furthermore, we conducted a comparative analysis of different prompting options within SAM and the fully-supervised nnU-Net. The results demonstrate the superior performance of our method compared to SAM employing only point prompts (approximately 50% improvement in IoU) and performs on-par with fully supervised methods whilst reducing the requirement of labeled data by at least an order of magnitude.
Serialized Output Training (SOT) has showcased state-of-the-art performance in multi-talker speech recognition by sequentially decoding the speech of individual speakers. To address the challenging label-permutation issue, prior methods have relied on either the Permutation Invariant Training (PIT) or the time-based First-In-First-Out (FIFO) rule. This study presents a model-based serialization strategy that incorporates an auxiliary module into the Attention Encoder-Decoder architecture, autonomously identifying the crucial factors to order the output sequence of the speech components in multi-talker speech. Experiments conducted on the LibriSpeech and LibriMix databases reveal that our approach significantly outperforms the PIT and FIFO baselines in both 2-mix and 3-mix scenarios. Further analysis shows that the serialization module identifies dominant speech components in a mixture by factors including loudness and gender, and orders speech components based on the dominance score.
Discretizing speech into tokens and generating them by a decoder-only model have been a promising direction for text-to-speech (TTS) and spoken language modeling (SLM). To shorten the sequence length of speech tokens, acoustic byte-pair encoding (BPE) has emerged in SLM that treats speech tokens from self-supervised semantic representations as characters to further compress the token sequence. But the gain in TTS has not been fully investigated, and the proper choice of acoustic BPE remains unclear. In this work, we conduct a comprehensive study on various settings of acoustic BPE to explore its effectiveness in decoder-only TTS models with semantic speech tokens. Experiments on LibriTTS verify that acoustic BPE uniformly increases the intelligibility and diversity of synthesized speech, while showing different features across BPE settings. Hence, acoustic BPE is a favorable tool for decoder-only TTS.
We contribute an unsupervised method that effectively learns from raw observation and disentangles its latent space into content and style representations. Unlike most disentanglement algorithms that rely on domain-specific labels and knowledge, our method is based on the insight of domain-general statistical differences between content and style -- content varies more among different fragments within a sample but maintains an invariant vocabulary across data samples, whereas style remains relatively invariant within a sample but exhibits more significant variation across different samples. We integrate such inductive bias into an encoder-decoder architecture and name our method after V3 (variance-versus-invariance). Experimental results show that V3 generalizes across two distinct domains in different modalities, music audio and images of written digits, successfully learning pitch-timbre and digit-color disentanglements, respectively. Also, the disentanglement robustness significantly outperforms baseline unsupervised methods and is even comparable to supervised counterparts. Furthermore, symbolic-level interpretability emerges in the learned codebook of content, forging a near one-to-one alignment between machine representation and human knowledge.
The equivalent nature of 3D coordinates has posed long term challenges in protein structure representation learning, alignment, and generation. Can we create a compact and invariant language that equivalently represents protein structures? Towards this goal, we propose FoldToken2 to transfer equivariant structures into discrete tokens, while maintaining the recoverability of the original structures. From FoldToken1 to FoldToken2, we improve three key components: (1) invariant structure encoder, (2) vector-quantized compressor, and (3) equivalent structure decoder. We evaluate FoldToken2 on the protein structure reconstruction task and show that it outperforms previous FoldToken1 by 20\% in TMScore and 81\% in RMSD. FoldToken2 probably be the first method that works well on both single-chain and multi-chain protein structures quantization. We believe that FoldToken2 will inspire further improvement in protein structure representation learning, structure alignment, and structure generation tasks.
While recent advances in Text-To-Speech synthesis have yielded remarkable improvements in generating high-quality speech, research on lightweight and fast models is limited. This paper introduces FLY-TTS, a new fast, lightweight and high-quality speech synthesis system based on VITS. Specifically, 1) We replace the decoder with ConvNeXt blocks that generate Fourier spectral coefficients followed by the inverse short-time Fourier transform to synthesize waveforms; 2) To compress the model size, we introduce the grouped parameter-sharing mechanism to the text encoder and flow-based model; 3) We further employ the large pre-trained WavLM model for adversarial training to improve synthesis quality. Experimental results show that our model achieves a real-time factor of 0.0139 on an Intel Core i9 CPU, 8.8x faster than the baseline (0.1221), with a 1.6x parameter compression. Objective and subjective evaluations indicate that FLY-TTS exhibits comparable speech quality to the strong baseline.
Tashkeel, or Arabic Text Diacritization (ATD), greatly enhances the comprehension of Arabic text by removing ambiguity and minimizing the risk of misinterpretations caused by its absence. It plays a crucial role in improving Arabic text processing, particularly in applications such as text-to-speech and machine translation. This paper introduces a new approach to training ATD models. First, we finetuned two transformers, encoder-only and encoder-decoder, that were initialized from a pretrained character-based BERT. Then, we applied the Noisy-Student approach to boost the performance of the best model. We evaluated our models alongside 11 commercial and open-source models using two manually labeled benchmark datasets: WikiNews and our CATT dataset. Our findings show that our top model surpasses all evaluated models by relative Diacritic Error Rates (DERs) of 30.83\% and 35.21\% on WikiNews and CATT, respectively, achieving state-of-the-art in ATD. In addition, we show that our model outperforms GPT-4-turbo on CATT dataset by a relative DER of 9.36\%. We open-source our CATT models and benchmark dataset for the research community\footnote{https://github.com/abjadai/catt}.
Large language models (LLMs), known for their exceptional reasoning capabilities, generalizability, and fluency across diverse domains, present a promising avenue for enhancing speech-related tasks. In this paper, we focus on integrating decoder-only LLMs to the task of speech-to-text translation (S2TT). We propose a decoder-only architecture that enables the LLM to directly consume the encoded speech representation and generate the text translation. Additionally, we investigate the effects of different parameter-efficient fine-tuning techniques and task formulation. Our model achieves state-of-the-art performance on CoVoST 2 and FLEURS among models trained without proprietary data. We also conduct analyses to validate the design choices of our proposed model and bring insights to the integration of LLMs to S2TT.