Reversible face anonymization, unlike traditional face pixelization, seeks to replace sensitive identity information in facial images with synthesized alternatives, preserving privacy without sacrificing image clarity. Traditional methods, such as encoder-decoder networks, often result in significant loss of facial details due to their limited learning capacity. Additionally, relying on latent manipulation in pre-trained GANs can lead to changes in ID-irrelevant attributes, adversely affecting data utility due to GAN inversion inaccuracies. This paper introduces G\textsuperscript{2}Face, which leverages both generative and geometric priors to enhance identity manipulation, achieving high-quality reversible face anonymization without compromising data utility. We utilize a 3D face model to extract geometric information from the input face, integrating it with a pre-trained GAN-based decoder. This synergy of generative and geometric priors allows the decoder to produce realistic anonymized faces with consistent geometry. Moreover, multi-scale facial features are extracted from the original face and combined with the decoder using our novel identity-aware feature fusion blocks (IFF). This integration enables precise blending of the generated facial patterns with the original ID-irrelevant features, resulting in accurate identity manipulation. Extensive experiments demonstrate that our method outperforms existing state-of-the-art techniques in face anonymization and recovery, while preserving high data utility. Code is available at https://github.com/Harxis/G2Face.
Federated Collaborative Filtering (FedCF) is an emerging field focused on developing a new recommendation framework with preserving privacy in a federated setting. Existing FedCF methods typically combine distributed Collaborative Filtering (CF) algorithms with privacy-preserving mechanisms, and then preserve personalized information into a user embedding vector. However, the user embedding is usually insufficient to preserve the rich information of the fine-grained personalization across heterogeneous clients. This paper proposes a novel personalized FedCF method by preserving users' personalized information into a latent variable and a neural model simultaneously. Specifically, we decompose the modeling of user knowledge into two encoders, each designed to capture shared knowledge and personalized knowledge separately. A personalized gating network is then applied to balance personalization and generalization between the global and local encoders. Moreover, to effectively train the proposed framework, we model the CF problem as a specialized Variational AutoEncoder (VAE) task by integrating user interaction vector reconstruction with missing value prediction. The decoder is trained to reconstruct the implicit feedback from items the user has interacted with, while also predicting items the user might be interested in but has not yet interacted with. Experimental results on benchmark datasets demonstrate that the proposed method outperforms other baseline methods, showcasing superior performance. Our code is available at https://github.com/mtics/FedDAE.
Facial wrinkle detection plays a crucial role in cosmetic dermatology. Precise manual segmentation of facial wrinkles is challenging and time-consuming, with inherent subjectivity leading to inconsistent results among graders. To address this issue, we propose two solutions. First, we build and release the first public facial wrinkle dataset, 'FFHQ-Wrinkle', an extension of the NVIDIA FFHQ dataset. It includes 1,000 images with human labels and 50,000 images with automatically generated weak labels. This dataset could serve as a foundation for the research community to develop advanced wrinkle detection algorithms. Second, we introduce a simple training strategy utilizing texture maps, applicable to various segmentation models, to detect wrinkles across the face. Our two-stage training strategy first pretrain models on a large dataset with weak labels (N=50k), or masked texture maps generated through computer vision techniques, without human intervention. We then finetune the models using human-labeled data (N=1k), which consists of manually labeled wrinkle masks. The network takes as input a combination of RGB and masked texture map of the image, comprising four channels, in finetuning. We effectively combine labels from multiple annotators to minimize subjectivity in manual labeling. Our strategies demonstrate improved segmentation performance in facial wrinkle segmentation both quantitatively and visually compared to existing pretraining methods. The dataset is available at https://github.com/labhai/ffhq-wrinkle-dataset.
We highlight a perhaps important but hitherto unobserved insight: The attention module in a transformer is a smoothed cubic spline. Viewed in this manner, this mysterious but critical component of a transformer becomes a natural development of an old notion deeply entrenched in classical approximation theory. More precisely, we show that with ReLU-activation, attention, masked attention, encoder-decoder attention are all cubic splines. As every component in a transformer is constructed out of compositions of various attention modules (= cubic splines) and feed forward neural networks (= linear splines), all its components -- encoder, decoder, and encoder-decoder blocks; multilayered encoders and decoders; the transformer itself -- are cubic or higher-order splines. If we assume the Pierce-Birkhoff conjecture, then the converse also holds, i.e., every spline is a ReLU-activated encoder. Since a spline is generally just $C^2$, one way to obtain a smoothed $C^\infty$-version is by replacing ReLU with a smooth activation; and if this activation is chosen to be SoftMax, we recover the original transformer as proposed by Vaswani et al. This insight sheds light on the nature of the transformer by casting it entirely in terms of splines, one of the best known and thoroughly understood objects in applied mathematics.
Online vectorized High-Definition (HD) map construction is crucial for subsequent prediction and planning tasks in autonomous driving. Following MapTR paradigm, recent works have made noteworthy achievements. However, reference points are randomly initialized in mainstream methods, leading to unstable matching between predictions and ground truth. To address this issue, we introduce PriorMapNet to enhance online vectorized HD map construction with priors. We propose the PPS-Decoder, which provides reference points with position and structure priors. Fitted from the map elements in the dataset, prior reference points lower the learning difficulty and achieve stable matching. Furthermore, we propose the PF-Encoder to enhance the image-to-BEV transformation with BEV feature priors. Besides, we propose the DMD cross-attention, which decouples cross-attention along multi-scale and multi-sample respectively to achieve efficiency. Our proposed PriorMapNet achieves state-of-the-art performance in the online vectorized HD map construction task on nuScenes and Argoverse2 datasets. The code will be released publicly soon.
Masked autoencoder has been widely explored in point cloud self-supervised learning, whereby the point cloud is generally divided into visible and masked parts. These methods typically include an encoder accepting visible patches (normalized) and corresponding patch centers (position) as input, with the decoder accepting the output of the encoder and the centers (position) of the masked parts to reconstruct each point in the masked patches. Then, the pre-trained encoders are used for downstream tasks. In this paper, we show a motivating empirical result that when directly feeding the centers of masked patches to the decoder without information from the encoder, it still reconstructs well. In other words, the centers of patches are important and the reconstruction objective does not necessarily rely on representations of the encoder, thus preventing the encoder from learning semantic representations. Based on this key observation, we propose a simple yet effective method, i.e., learning to Predict Centers for Point Masked AutoEncoders (PCP-MAE) which guides the model to learn to predict the significant centers and use the predicted centers to replace the directly provided centers. Specifically, we propose a Predicting Center Module (PCM) that shares parameters with the original encoder with extra cross-attention to predict centers. Our method is of high pre-training efficiency compared to other alternatives and achieves great improvement over Point-MAE, particularly surpassing it by 5.50% on OBJ-BG, 6.03% on OBJ-ONLY, and 5.17% on PB-T50-RS for 3D object classification on the ScanObjectNN dataset. The code is available at https://github.com/aHapBean/PCP-MAE.
The Segment Anything Model (SAM), a foundational model designed for promptable segmentation tasks, demonstrates exceptional generalization capabilities, making it highly promising for natural scene image segmentation. However, SAM's lack of pretraining on massive remote sensing images and its interactive structure limit its automatic mask prediction capabilities. In this paper, a Multi-Cognitive SAM-Based Instance Segmentation Model (MC-SAM SEG) is introduced to employ SAM on remote sensing domain. The SAM-Mona encoder utilizing the Multi-cognitive Visual Adapter (Mona) is conducted to facilitate SAM's transfer learning in remote sensing applications. The proposed method named MC-SAM SEG extracts high-quality features by fine-tuning the SAM-Mona encoder along with a feature aggregator. Subsequently, a pixel decoder and transformer decoder are designed for prompt-free mask generation and instance classification. The comprehensive experiments are conducted on the HRSID and WHU datasets for instance segmentation tasks on Synthetic Aperture Radar (SAR) images and optical remote sensing images respectively. The evaluation results indicate the proposed method surpasses other deep learning algorithms and verify its effectiveness and generalization.
We develop a large language model (LLM) based automatic speech recognition (ASR) system that can be contextualized by providing keywords as prior information in text prompts. We adopt decoder-only architecture and use our in-house LLM, PLaMo-100B, pre-trained from scratch using datasets dominated by Japanese and English texts as the decoder. We adopt a pre-trained Whisper encoder as an audio encoder, and the audio embeddings from the audio encoder are projected to the text embedding space by an adapter layer and concatenated with text embeddings converted from text prompts to form inputs to the decoder. By providing keywords as prior information in the text prompts, we can contextualize our LLM-based ASR system without modifying the model architecture to transcribe ambiguous words in the input audio accurately. Experimental results demonstrate that providing keywords to the decoder can significantly improve the recognition performance of rare and ambiguous words.
Full waveform inversion (FWI) plays a crucial role in the field of geophysics. There has been lots of research about applying deep learning (DL) methods to FWI. The success of DL-FWI relies significantly on the quantity and diversity of the datasets. Nevertheless, existing FWI datasets, like OpenFWI, where sources have fixed locations or identical frequencies, provide limited information and do not represent the complex real-world scene. For instance, low frequencies help in resolving larger-scale structures. High frequencies allow for a more detailed subsurface features. %A single source frequency is insufficient to describe subsurface structural properties. We consider that simultaneously using sources with different frequencies, instead of performing inversion using low frequencies data and then gradually introducing higher frequencies data, has rationale and potential advantages. Hence, we develop three enhanced datasets based on OpenFWI where each source have varying locations, frequencies or both. Moreover, we propose a novel deep operator network (DeepONet) architecture Inversion-DeepONet for FWI. We utilize convolutional neural network (CNN) to extract the features from seismic data in branch net. Source parameters, such as locations and frequencies, are fed to trunk net. Then another CNN is employed as the decoder of DeepONet to reconstruct the velocity models more effectively. Through experiments, we confirm the superior performance on accuracy and generalization ability of our network, compared with existing data-driven FWI methods.
Recent advances in camera design and imaging technology have enabled the capture of high-quality images using smartphones. However, due to the limited dynamic range of digital cameras, the quality of photographs captured in environments with highly imbalanced lighting often results in poor-quality images. To address this issue, most devices capture multi-exposure frames and then use some multi-exposure fusion method to merge those frames into a final fused image. Nevertheless, most traditional and current deep learning approaches are unsuitable for real-time applications on mobile devices due to their heavy computational and memory requirements. We propose a new method for multi-exposure fusion based on an encoder-decoder deep learning architecture with efficient building blocks tailored for mobile devices. This efficient design makes our model capable of processing 4K resolution images in less than 2 seconds on mid-range smartphones. Our method outperforms state-of-the-art techniques regarding full-reference quality measures and computational efficiency (runtime and memory usage), making it ideal for real-time applications on hardware-constrained devices. Our code is available at: https://github.com/LucasKirsten/MobileMEF.
3D dense captioning is a task to localize objects in a 3D scene and generate descriptive sentences for each object. Recent approaches in 3D dense captioning have adopted transformer encoder-decoder frameworks from object detection to build an end-to-end pipeline without hand-crafted components. However, these approaches struggle with contradicting objectives where a single query attention has to simultaneously view both the tightly localized object regions and contextual environment. To overcome this challenge, we introduce SIA (See-It-All), a transformer pipeline that engages in 3D dense captioning with a novel paradigm called late aggregation. SIA simultaneously decodes two sets of queries-context query and instance query. The instance query focuses on localization and object attribute descriptions, while the context query versatilely captures the region-of-interest of relationships between multiple objects or with the global scene, then aggregated afterwards (i.e., late aggregation) via simple distance-based measures. To further enhance the quality of contextualized caption generation, we design a novel aggregator to generate a fully informed caption based on the surrounding context, the global environment, and object instances. Extensive experiments on two of the most widely-used 3D dense captioning datasets demonstrate that our proposed method achieves a significant improvement over prior methods.
Latent space geometry provides a rigorous and empirically valuable framework for interacting with the latent variables of deep generative models. This approach reinterprets Euclidean latent spaces as Riemannian through a pull-back metric, allowing for a standard differential geometric analysis of the latent space. Unfortunately, data manifolds are generally compact and easily disconnected or filled with holes, suggesting a topological mismatch to the Euclidean latent space. The most established solution to this mismatch is to let uncertainty be a proxy for topology, but in neural network models, this is often realized through crude heuristics that lack principle and generally do not scale to high-dimensional representations. We propose using ensembles of decoders to capture model uncertainty and show how to easily compute geodesics on the associated expected manifold. Empirically, we find this simple and reliable, thereby coming one step closer to easy-to-use latent geometries.
Beyond the Transformer, it is important to explore how to exploit the capacity of the MetaFormer, an architecture that is fundamental to the performance improvements of the Transformer. Previous studies have exploited it only for the backbone network. Unlike previous studies, we explore the capacity of the Metaformer architecture more extensively in the semantic segmentation task. We propose a powerful semantic segmentation network, MetaSeg, which leverages the Metaformer architecture from the backbone to the decoder. Our MetaSeg shows that the MetaFormer architecture plays a significant role in capturing the useful contexts for the decoder as well as for the backbone. In addition, recent segmentation methods have shown that using a CNN-based backbone for extracting the spatial information and a decoder for extracting the global information is more effective than using a transformer-based backbone with a CNN-based decoder. This motivates us to adopt the CNN-based backbone using the MetaFormer block and design our MetaFormer-based decoder, which consists of a novel self-attention module to capture the global contexts. To consider both the global contexts extraction and the computational efficiency of the self-attention for semantic segmentation, we propose a Channel Reduction Attention (CRA) module that reduces the channel dimension of the query and key into the one dimension. In this way, our proposed MetaSeg outperforms the previous state-of-the-art methods with more efficient computational costs on popular semantic segmentation and a medical image segmentation benchmark, including ADE20K, Cityscapes, COCO-stuff, and Synapse. The code is available at https://github.com/hyunwoo137/MetaSeg.
This paper describes CMU's submission to the IWSLT 2024 Simultaneous Speech Translation (SST) task for translating English speech to German text in a streaming manner. Our end-to-end speech-to-text (ST) system integrates the WavLM speech encoder, a modality adapter, and the Llama2-7B-Base model as the decoder. We employ a two-stage training approach: initially, we align the representations of speech and text, followed by full fine-tuning. Both stages are trained on MuST-c v2 data with cross-entropy loss. We adapt our offline ST model for SST using a simple fixed hold-n policy. Experiments show that our model obtains an offline BLEU score of 31.1 and a BLEU score of 29.5 under 2 seconds latency on the MuST-C-v2 tst-COMMON.
Semantic communication (SemComm) has emerged as new paradigm shifts.Most existing SemComm systems transmit continuously distributed signals in analog fashion.However, the analog paradigm is not compatible with current digital communication frameworks. In this paper, we propose an alternating multi-phase training strategy (AMP) to enable the joint training of the networks in the encoder and decoder through non-differentiable digital processes. AMP contains three training phases, aiming at feature extraction (FE), robustness enhancement (RE), and training-testing alignment (TTA), respectively. AMP contains three training phases, aiming at feature extraction (FE), robustness enhancement (RE), and training-testing alignment (TTA), respectively. In particular, in the FE stage, we learn the representation ability of semantic information by end-to-end training the encoder and decoder in an analog manner. When we take digital communication into consideration, the domain shift between digital and analog demands the fine-tuning for encoder and decoder. To cope with joint training process within the non-differentiable digital processes, we propose the alternation between updating the decoder individually and jointly training the codec in RE phase. To boost robustness further, we investigate a mask-attack (MATK) in RE to simulate an evident and severe bit-flipping effect in a differentiable manner. To address the training-testing inconsistency introduced by MATK, we employ an additional TTA phase, fine-tuning the decoder without MATK. Combining with AMP and an information restoration network, we propose a digital SemComm system for image transmission, named AMP-SC. Comparing with the representative benchmark, AMP-SC achieves $0.82 \sim 1.65$dB higher average reconstruction performance among various representative datasets at different scales and a wide range of signal-to-noise ratio.
We propose a generative adversarial network (GAN) based deep learning method that serves the dual role of both identification and mitigation of cyber-attacks in wide-area damping control loops of power systems. Two specific types of attacks considered are false data injection and denial-of-service (DoS). Unlike existing methods, which are either model-based or model-free and yet require two separate learning modules for detection and mitigation leading to longer response times before clearing an attack, our deep learner incorporate both goals within the same integrated framework. A Long Short-Term Memory (LSTM) encoder-decoder based GAN is proposed that captures the temporal dynamics of the power system significantly more accurately than fully-connected GANs, thereby providing better accuracy and faster response for both goals. The method is validated using the IEEE 68-bus power system model.
Transcribing the speech of multiple overlapping speakers typically requires separating the audio into multiple streams and recognizing each one independently. More recent work jointly separates and transcribes, but requires a separate decoding component for each speaker. We propose the TOGGL model to simultaneously transcribe the speech of multiple speakers. The TOGGL model uses special output tokens to attribute the speech to each speaker with only a single decoder. Our approach generalizes beyond two speakers, even when trained only on two-speaker data. We demonstrate superior performance compared to competing approaches on a conversational speech dataset. Our approach also improves performance on single-speaker audio.
Vision-Language Models (VLMs) have emerged as general purpose tools for addressing a variety of complex computer vision problems. Such models have been shown to be highly capable, but, at the same time, also lacking some basic visual understanding skills. In this paper, we set out to understand the limitations of SoTA VLMs on fundamental visual tasks: object classification, understanding spatial arrangement, and ability to delineate individual object instances (through counting), by constructing a series of tests that probe which components of design, specifically, maybe lacking. Importantly, we go significantly beyond the current benchmarks, that simply measure final performance of VLM, by also comparing and contrasting it to performance of probes trained directly on features obtained from visual encoder (image embeddings), as well as intermediate vision-language projection used to bridge image-encoder and LLM-decoder ouput in many SoTA models (e.g., LLaVA, BLIP, InstructBLIP). In doing so, we uncover nascent shortcomings in VLMs response and make a number of important observations which could help train and develop more effective VLM models in future.
Tables are crucial containers of information, but understanding their meaning may be challenging. Over the years, there has been a surge in interest in data-driven approaches based on deep learning that have increasingly been combined with heuristic-based ones. In the last period, the advent of \acf{llms} has led to a new category of approaches for table annotation. However, these approaches have not been consistently evaluated on a common ground, making evaluation and comparison difficult. This work proposes an extensive evaluation of four STI SOTA approaches: Alligator (formerly s-elbat), Dagobah, TURL, and TableLlama; the first two belong to the family of heuristic-based algorithms, while the others are respectively encoder-only and decoder-only Large Language Models (LLMs). We also include in the evaluation both GPT-4o and GPT-4o-mini, since they excel in various public benchmarks. The primary objective is to measure the ability of these approaches to solve the entity disambiguation task with respect to both the performance achieved on a common-ground evaluation setting and the computational and cost requirements involved, with the ultimate aim of charting new research paths in the field.
Open Domain Question Answering (ODQA) has been advancing rapidly in recent times, driven by significant developments in dense passage retrieval and pretrained language models. Current models typically incorporate the FiD framework, which is composed by a neural retriever alongside an encoder-decoder neural reader. In the answer generation process, the retriever will retrieve numerous passages (around 100 for instance), each of which is then individually encoded by the encoder. Subsequently, the decoder makes predictions based on these encoded passages. Nevertheless, this framework can be relatively time-consuming, particularly due to the extensive length of the gathered passages. To address this, we introduce FastFiD in this paper, a novel approach that executes sentence selection on the encoded passages. This aids in retaining valuable sentences while reducing the context length required for generating answers. Experiments on three commonly used datasets (Natural Questions, TriviaQA and ASQA) demonstrate that our method can enhance the inference speed by 2.3X-5.7X, while simultaneously maintaining the model's performance. Moreover, an in-depth analysis of the model's attention reveals that the selected sentences indeed hold a substantial contribution towards the final answer. The codes are publicly available at https://github.com/thunlp/FastFiD.