PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction

In large vision-language models (LVLMs), images serve as inputs that carry a wealth of information. As the idiom "A picture is worth a thousand words" implies, representing a single image in current LVLMs can require hundreds or even thousands of tokens. This results in significant computational costs, which grow quadratically as input image resolution increases, thereby severely impacting the efficiency of both training and inference. Previous approaches have attempted to reduce the number of image tokens either before or within the early layers of LVLMs. However, these strategies inevitably result in the loss of crucial image information, ultimately diminishing model performance. To address this challenge, we conduct an empirical study revealing that all visual tokens are necessary for LVLMs in the shallow layers, and token redundancy progressively increases in the deeper layers of the model. To this end, we propose PyramidDrop, a visual redundancy reduction strategy for LVLMs to boost their efficiency in both training and inference with neglectable performance loss. Specifically, we partition the LVLM into several stages and drop part of the image tokens at the end of each stage with a pre-defined ratio, creating pyramid-like visual tokens across model layers. The dropping is based on a lightweight similarity calculation with a negligible time overhead. Extensive experiments demonstrate that PyramidDrop can achieve a 40% training time and 55% inference FLOPs acceleration of LLaVA-NeXT with comparable performance. Besides, the PyramidDrop could also serve as a plug-and-play strategy for inference acceleration without training, with better performance and lower inference cost than counterparts. We hope that the insights and approach introduced by PyramidDrop will inspire future research to further investigate the role of image tokens in LVLMs.

Coarse-to-Fine Highlighting: Reducing Knowledge Hallucination in Large Language Models

Generation of plausible but incorrect factual information, often termed hallucination, has attracted significant research interest. Retrieval-augmented language model (RALM) -- which enhances models with up-to-date knowledge -- emerges as a promising method to reduce hallucination. However, existing RALMs may instead exacerbate hallucination when retrieving lengthy contexts. To address this challenge, we propose COFT, a novel \textbf{CO}arse-to-\textbf{F}ine highligh\textbf{T}ing method to focus on different granularity-level key texts, thereby avoiding getting lost in lengthy contexts. Specifically, COFT consists of three components: \textit{recaller}, \textit{scorer}, and \textit{selector}. First, \textit{recaller} applies a knowledge graph to extract potential key entities in a given context. Second, \textit{scorer} measures the importance of each entity by calculating its contextual weight. Finally, \textit{selector} selects high contextual weight entities with a dynamic threshold algorithm and highlights the corresponding paragraphs, sentences, or words in a coarse-to-fine manner. Extensive experiments on the knowledge hallucination benchmark demonstrate the effectiveness of COFT, leading to a superior performance over $30\%$ in the F1 score metric. Moreover, COFT also exhibits remarkable versatility across various long-form tasks, such as reading comprehension and question answering.

D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement

We introduce D-FINE, a powerful real-time object detector that achieves outstanding localization precision by redefining the bounding box regression task in DETR models. D-FINE comprises two key components: Fine-grained Distribution Refinement (FDR) and Global Optimal Localization Self-Distillation (GO-LSD). FDR transforms the regression process from predicting fixed coordinates to iteratively refining probability distributions, providing a fine-grained intermediate representation that significantly enhances localization accuracy. GO-LSD is a bidirectional optimization strategy that transfers localization knowledge from refined distributions to shallower layers through self-distillation, while also simplifying the residual prediction tasks for deeper layers. Additionally, D-FINE incorporates lightweight optimizations in computationally intensive modules and operations, achieving a better balance between speed and accuracy. Specifically, D-FINE-L / X achieves 54.0% / 55.8% AP on the COCO dataset at 124 / 78 FPS on an NVIDIA T4 GPU. When pretrained on Objects365, D-FINE-L / X attains 57.1% / 59.3% AP, surpassing all existing real-time detectors. Furthermore, our method significantly enhances the performance of a wide range of DETR models by up to 5.3% AP with negligible extra parameters and training costs. Our code and pretrained models: https://github.com/Peterande/D-FINE.

USTC-TD: A Test Dataset and Benchmark for Image and Video Coding in 2020s

Image/video coding has been a remarkable research area for both academia and industry for many years. Testing datasets, especially high-quality image/video datasets are desirable for the justified evaluation of coding-related research, practical applications, and standardization activities. We put forward a test dataset namely USTC-TD, which has been successfully adopted in the practical end-to-end image/video coding challenge of the IEEE International Conference on Visual Communications and Image Processing in 2022 and 2023. USTC-TD contains 40 images at 4K spatial resolution and 10 video sequences at 1080p spatial resolution, featuring various content due to the diverse environmental factors (scene type, texture, motion, view) and the designed imaging factors (illumination, shadow, lens). We quantitatively evaluate USTC-TD on different image/video features (spatial, temporal, color, lightness), and compare it with the previous image/video test datasets, which verifies the wider coverage and more diversity of the proposed dataset. We also evaluate both classic standardized and recent learned image/video coding schemes on USTC-TD with PSNR and MS-SSIM, and provide an extensive benchmark for the evaluated schemes. Based on the characteristics and specific design of the proposed test dataset, we analyze the benchmark performance and shed light on the future research and development of image/video coding. All the data are released online: https://esakak.github.io/USTC-TD.

ClickAttention: Click Region Similarity Guided Interactive Segmentation

Interactive segmentation algorithms based on click points have garnered significant attention from researchers in recent years. However, existing studies typically use sparse click maps as model inputs to segment specific target objects, which primarily affect local regions and have limited abilities to focus on the whole target object, leading to increased times of clicks. In addition, most existing algorithms can not balance well between high performance and efficiency. To address this issue, we propose a click attention algorithm that expands the influence range of positive clicks based on the similarity between positively-clicked regions and the whole input. We also propose a discriminative affinity loss to reduce the attention coupling between positive and negative click regions to avoid an accuracy decrease caused by mutual interference between positive and negative clicks. Extensive experiments demonstrate that our approach is superior to existing methods and achieves cutting-edge performance in fewer parameters. An interactive demo and all reproducible codes will be released at https://github.com/hahamyt/ClickAttention.

NVC-1B: A Large Neural Video Coding Model

The emerging large models have achieved notable progress in the fields of natural language processing and computer vision. However, large models for neural video coding are still unexplored. In this paper, we try to explore how to build a large neural video coding model. Based on a small baseline model, we gradually scale up the model sizes of its different coding parts, including the motion encoder-decoder, motion entropy model, contextual encoder-decoder, contextual entropy model, and temporal context mining module, and analyze the influence of model sizes on video compression performance. Then, we explore to use different architectures, including CNN, mixed CNN-Transformer, and Transformer architectures, to implement the neural video coding model and analyze the influence of model architectures on video compression performance. Based on our exploration results, we design the first neural video coding model with more than 1 billion parameters -- NVC-1B. Experimental results show that our proposed large model achieves a significant video compression performance improvement over the small baseline model, and represents the state-of-the-art compression efficiency. We anticipate large models may bring up the video coding technologies to the next level.

Uniformly Accelerated Motion Model for Inter Prediction

Inter prediction is a key technology to reduce the temporal redundancy in video coding. In natural videos, there are usually multiple moving objects with variable velocity, resulting in complex motion fields that are difficult to represent compactly. In Versatile Video Coding (VVC), existing inter prediction methods usually assume uniform speed motion between consecutive frames and use the linear models for motion estimation (ME) and motion compensation (MC), which may not well handle the complex motion fields in the real world. To address these issues, we introduce a uniformly accelerated motion model (UAMM) to exploit motion-related elements (velocity, acceleration) of moving objects between the video frames, and further combine them to assist the inter prediction methods to handle the variable motion in the temporal domain. Specifically, first, the theory of UAMM is mentioned. Second, based on that, we propose the UAMM-based parameter derivation and extrapolation schemes in the coding process. Third, we integrate the UAMM into existing inter prediction modes (Merge, MMVD, CIIP) to achieve higher prediction accuracy. The proposed method is implemented into the VVC reference software, VTM version 12.0. Experimental results show that the proposed method achieves up to 0.38% and on average 0.13% BD-rate reduction compared to the VTM anchor, under the Low-delay P configuration, with a slight increase of time complexity on the encoding/decoding side.

In-Loop Filtering via Trained Look-Up Tables

In-loop filtering (ILF) is a key technology for removing the artifacts in image/video coding standards. Recently, neural network-based in-loop filtering methods achieve remarkable coding gains beyond the capability of advanced video coding standards, which becomes a powerful coding tool candidate for future video coding standards. However, the utilization of deep neural networks brings heavy time and computational complexity, and high demands of high-performance hardware, which is challenging to apply to the general uses of coding scene. To address this limitation, inspired by explorations in image restoration, we propose an efficient and practical in-loop filtering scheme by adopting the Look-up Table (LUT). We train the DNN of in-loop filtering within a fixed filtering reference range, and cache the output values of the DNN into a LUT via traversing all possible inputs. At testing time in the coding process, the filtered pixel is generated by locating input pixels (to-be-filtered pixel with reference pixels) and interpolating cached filtered pixel values. To further enable the large filtering reference range with the limited storage cost of LUT, we introduce the enhanced indexing mechanism in the filtering process, and clipping/finetuning mechanism in the training. The proposed method is implemented into the Versatile Video Coding (VVC) reference software, VTM-11.0. Experimental results show that the ultrafast, very fast, and fast mode of the proposed method achieves on average 0.13%/0.34%/0.51%, and 0.10%/0.27%/0.39% BD-rate reduction, under the all intra (AI) and random access (RA) configurations. Especially, our method has friendly time and computational complexity, only 101%/102%-104%/108% time increase with 0.13-0.93 kMACs/pixel, and only 164-1148 KB storage cost for a single model. Our solution may shed light on the journey of practical neural network-based coding tool evolution.

G-Transformer: Counterfactual Outcome Prediction under Dynamic and Time-varying Treatment Regimes

In the context of medical decision making, counterfactual prediction enables clinicians to predict treatment outcomes of interest under alternative courses of therapeutic actions given observed patient history. Prior machine learning approaches for counterfactual predictions under time-varying treatments focus on static time-varying treatment regimes where treatments do not depend on previous covariate history. In this work, we present G-Transformer, a Transformer-based framework supporting g-computation for counterfactual prediction under dynamic and time-varying treatment strategies. G-Transfomer captures complex, long-range dependencies in time-varying covariates using a Transformer architecture. G-Transformer estimates the conditional distribution of relevant covariates given covariate and treatment history at each time point using an encoder architecture, then produces Monte Carlo estimates of counterfactual outcomes by simulating forward patient trajectories under treatment strategies of interest. We evaluate G-Transformer extensively using two simulated longitudinal datasets from mechanistic models, and a real-world sepsis ICU dataset from MIMIC-IV. G-Transformer outperforms both classical and state-of-the-art counterfactual prediction models in these settings. To the best of our knowledge, this is the first Transformer-based architecture for counterfactual outcome prediction under dynamic and time-varying treatment strategies. Code will be released upon publication of the paper.

Inference Attacks in Machine Learning as a Service: A Taxonomy, Review, and Promising Directions

The prosperity of machine learning has also brought people's concerns about data privacy. Among them, inference attacks can implement privacy breaches in various MLaaS scenarios and model training/prediction phases. Specifically, inference attacks can perform privacy inference on undisclosed target training sets based on outputs of the target model, including but not limited to statistics, membership, semantics, data representation, etc. For instance, infer whether the target data has the characteristics of AIDS. In addition, the rapid development of the machine learning community in recent years, especially the surge of model types and application scenarios, has further stimulated the inference attacks' research. Thus, studying inference attacks and analyzing them in depth is urgent and significant. However, there is still a gap in the systematic discussion of inference attacks from taxonomy, global perspective, attack, and defense perspectives. This survey provides an in-depth and comprehensive inference of attacks and corresponding countermeasures in ML-as-a-service based on taxonomy and the latest researches. Without compromising researchers' intuition, we first propose the 3MP taxonomy based on the community research status, trying to normalize the confusing naming system of inference attacks. Also, we analyze the pros and cons of each type of inference attack, their workflow, countermeasure, and how they interact with other attacks. In the end, we point out several promising directions for researchers from a more comprehensive and novel perspective.