Zhang Zhang

Ziheng Jiang, Haibin Lin, Yinmin Zhong et al.

We present the design, implementation and engineering experience in building and deploying MegaScale, a production system for training large language models (LLMs) at the scale of more than 10,000 GPUs. Training LLMs at this scale brings unprecedented challenges to training efficiency and stability. We take a full-stack approach that co-designs the algorithmic and system components across model block and optimizer design, computation and communication overlapping, operator optimization, data pipeline, and network performance tuning. Maintaining high efficiency throughout the training process (i.e., stability) is an important consideration in production given the long extent of LLM training jobs. Many hard stability issues only emerge at large scale, and in-depth observability is the key to address them. We develop a set of diagnosis tools to monitor system components and events deep in the stack, identify root causes, and derive effective techniques to achieve fault tolerance and mitigate stragglers. MegaScale achieves 55.2% Model FLOPs Utilization (MFU) when training a 175B LLM model on 12,288 GPUs, improving the MFU by 1.34x compared to Megatron-LM. We share our operational experience in identifying and fixing failures and stragglers. We hope by articulating the problems and sharing our experience from a systems perspective, this work can inspire future LLM systems research.

Tianze Wang, Zhang Zhang, Chao Sun

The classification of the type of road surface (RSC) aims to utilize pavement features to identify the roughness, wet and dry conditions, and material information of the road surface. Due to its ability to effectively enhance road safety and traffic management, it has received widespread attention in recent years. In autonomous driving, accurate RSC allows vehicles to better understand the road environment, adjust driving strategies, and ensure a safer and more efficient driving experience. For a long time, vision-based RSC has been favored. However, existing visual classification methods have overlooked the exploration of fine-grained classification of pavement types (such as similar pavement textures). In this work, we propose a pure vision-based fine-grained RSC method for autonomous driving scenarios, which fuses local and global feature information through the stacking of convolutional and transformer modules. We further explore the stacking strategies of local and global feature extraction modules to find the optimal feature extraction strategy. In addition, since fine-grained tasks also face the challenge of relatively large intra-class differences and relatively small inter-class differences, we propose a Foreground-Background Module (FBM) that effectively extracts fine-grained context features of the pavement, enhancing the classification ability for complex pavements. Experiments conducted on a large-scale pavement dataset containing one million samples and a simplified dataset reorganized from this dataset achieved Top-1 classification accuracies of 92.52% and 96.50%, respectively, improving by 5.69% to 12.84% compared to SOTA methods. These results demonstrate that RoadFormer outperforms existing methods in RSC tasks, providing significant progress in improving the reliability of pavement perception in autonomous driving systems.

Wenbo Tian, Ruting Lin, Hongxian Zheng et al.

Existing intelligent sports analysis systems mainly focus on "scoring and visualization," often lacking automatic performance diagnosis and interpretable training guidance. Recent advances in Large Language Models (LLMs) and motion analysis techniques provide new opportunities to address the above limitations. In this paper, we propose SportsGPT, an LLM-driven framework for interpretable sports motion assessment and training guidance, which establishes a closed loop from motion time-series input to professional training guidance. First, given a set of high-quality target models, we introduce MotionDTW, a two-stage time series alignment algorithm designed for accurate keyframe extraction from skeleton-based motion sequences. Subsequently, we design a Knowledge-based Interpretable Sports Motion Assessment Model (KISMAM) to obtain a set of interpretable assessment metrics (e.g., insufficient extension) by contrasting the keyframes with the target models. Finally, we propose SportsRAG, a RAG-based training guidance model built upon Qwen3. Leveraging a 6B-token knowledge base, it prompts the LLM to generate professional training guidance by retrieving domain-specific QA pairs. Experimental results demonstrate that MotionDTW significantly outperforms traditional methods with lower temporal error and higher IoU scores. Furthermore, ablation studies validate the KISMAM and SportsRAG, confirming that SportsGPT surpasses general LLMs in diagnostic accuracy and professionalism.

Jiayao Tan, Fan Lyu, Chenggong Ni et al.

Continual Test-Time Adaptation (CTTA) aims to adapt a pre-trained model to a sequence of target domains during the test phase without accessing the source data. To adapt to unlabeled data from unknown domains, existing methods rely on constructing pseudo-labels for all samples and updating the model through self-training. However, these pseudo-labels often involve noise, leading to insufficient adaptation. To improve the quality of pseudo-labels, we propose a pseudo-label selection method for CTTA, called Pseudo Labeling Filter (PLF). The key idea of PLF is to keep selecting appropriate thresholds for pseudo-labels and identify reliable ones for self-training. Specifically, we present three principles for setting thresholds during continuous domain learning, including initialization, growth and diversity. Based on these principles, we design Self-Adaptive Thresholding to filter pseudo-labels. Additionally, we introduce a Class Prior Alignment (CPA) method to encourage the model to make diverse predictions for unknown domain samples. Through extensive experiments, PLF outperforms current state-of-the-art methods, proving its effectiveness in CTTA.

Yi-Fan Zhang, Zhang Zhang, Da Li et al.

Generalizable person Re-Identification (ReID) has attracted growing attention in recent computer vision community. In this work, we construct a structural causal model among identity labels, identity-specific factors (clothes/shoes color etc), and domain-specific factors (background, viewpoints etc). According to the causal analysis, we propose a novel Domain Invariant Representation Learning for generalizable person Re-Identification (DIR-ReID) framework. Specifically, we first propose to disentangle the identity-specific and domain-specific feature spaces, based on which we propose an effective algorithmic implementation for backdoor adjustment, essentially serving as a causal intervention towards the SCM. Extensive experiments have been conducted, showing that DIR-ReID outperforms state-of-the-art methods on large-scale domain generalization ReID benchmarks.

Yi-Fan Zhang, Weiqiang Ren, Zhang Zhang et al.

In object detection, bounding box regression (BBR) is a crucial step that determines the object localization performance. However, we find that most previous loss functions for BBR have two main drawbacks: (i) Both $\ell_n$-norm and IOU-based loss functions are inefficient to depict the objective of BBR, which leads to slow convergence and inaccurate regression results. (ii) Most of the loss functions ignore the imbalance problem in BBR that the large number of anchor boxes which have small overlaps with the target boxes contribute most to the optimization of BBR. To mitigate the adverse effects caused thereby, we perform thorough studies to exploit the potential of BBR losses in this paper. Firstly, an Efficient Intersection over Union (EIOU) loss is proposed, which explicitly measures the discrepancies of three geometric factors in BBR, i.e., the overlap area, the central point and the side length. After that, we state the Effective Example Mining (EEM) problem and propose a regression version of focal loss to make the regression process focus on high-quality anchor boxes. Finally, the above two parts are combined to obtain a new loss function, namely Focal-EIOU loss. Extensive experiments on both synthetic and real datasets are performed. Notable superiorities on both the convergence speed and the localization accuracy can be achieved over other BBR losses.

Kai Yu, Yang Zhou, Da Li et al.

Visual surveillance systems have become one of the largest data sources of Big Visual Data in real world. However, existing systems for video analysis still lack the ability to handle the problems of scalability, expansibility and error-prone, though great advances have been achieved in a number of visual recognition tasks and surveillance applications, e.g., pedestrian/vehicle detection, people/vehicle counting. Moreover, few algorithms explore the specific values/characteristics in large-scale surveillance videos. To address these problems in large-scale video analysis, we develop a scalable video parsing and evaluation platform through combining some advanced techniques for Big Data processing, including Spark Streaming, Kafka and Hadoop Distributed Filesystem (HDFS). Also, a Web User Interface is designed in the system, to collect users' degrees of satisfaction on the recognition tasks so as to evaluate the performance of the whole system. Furthermore, the highly extensible platform running on the long-term surveillance videos makes it possible to develop more intelligent incremental algorithms to enhance the performance of various visual recognition tasks.