Baicheng Bian

Baicheng Bian

Auto-generated profile, based on publicly available information

Baicheng Bian

University of Shanghai for Science and Technology

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
29 November 2025 in Robotics and Computer-Integrated Manufacturing
Verified
Journal Article
Closed access
BBBaicheng Bian
Baicheng Bian
Long ChenLong Chen
Long Chen
ZHZongwang Han
Zongwang Han
SWShiqing Wu
Shiqing Wu
WLWeiDong Li
WeiDong Li
HCHongguang Chen
Hongguang Chen

1 May 2026 in Engineering Research Express
Verified
Journal Article
Closed access
Long ChenLong Chen
Long Chen
ZWZixiang Wang
Zixiang Wang
BBBaicheng Bian
Baicheng Bian
Existing semi-supervised semantic segmentation methods often use a single threshold to filter pseudo-labels, this leads to the abandonment of many pixels with slightly low but still reliable confidence. Meanwhile, the pseudo-labels generated by the model itself inevitably contain errors, these errors accumulate gradually during training, causing the model to solidify in the wrong direction. To improve the utilization of unlabeled data and solve the confirmation bias problem in model self-training—thus optimizing semantic segmentation accuracy—this paper proposes a semi-supervised semantic segmentation method (called MvpMatch) based on a hierarchical threshold strategy and historical model ensemble prediction. The hierarchical threshold strategy constructs multi-level threshold intervals by introducing a threshold relaxation factor. It assigns differentiated loss weights to pseudo-labels in different intervals to achieve efficient use of pixel information. The historical model ensemble prediction dynamically weights and aggregates predictions from multiple historical models. This ‘dilutes’ a small number of errors from a single historical model with correct predictions from other models, ensuring higher reliability of the aggregated pseudo-labels. The proposed method is validated on two datasets: Pascal and Tugiao. Under different label ratios, MvpMatch significantly outperforms the baseline model UniMatch V2. Specifically, under low-label ratio settings, the mIoU of the Pascal dataset is improved by up to 1.7 percentage points, and the Tugiao dataset achieves the same level of improvement. Experimental results show that MvpMatch has significant advantages in segmentation accuracy and resisting the accumulation of pseudo-label errors, especially in scenarios with scarce labeled data.

30 October 2026 in Electronics
Verified
Journal Article
Open access
PDF file
ZHZong wang Han
Zong wang Han
Long ChenLong Chen
Long Chen
BBBaicheng Bian
Baicheng Bian
SWShiqing Wu
Shiqing Wu
The temporal-resolution gap between low-frame-rate RGB-D observations and fast object motion poses a fundamental challenge to long-term 6D object tracking. Under large inter-frame translations or rotations, the true pose may fall outside the basin of convergence of local RGB-D optimization initialized from the previous-frame estimate. We present EventProp6D, a reconstruction-free framework that uses asynchronous events to propagate a motion-conditioned pose-candidate space between consecutive RGB-D observations. Our key insight is that events need not independently recover a complete 6D pose. Instead, they provide high-temporal-resolution motion constraints that reshape the candidate space at the next RGB-D timestamp, extending it toward pose regions that local optimization cannot readily reach under fast motion. EventProp6D combines geometry-guided multi-scale translation propagation, which transports a CAD-anchored event support region while compensating for rotation-induced deformation, with motion-consistent SO(3) recovery, which preserves coherent angular motion and generates guarded hypotheses for abrupt rotations. The resulting SE(3) candidates are refined and selected using the current RGB-D observation. Experiments on the Event6D and EventHO3D datasets demonstrate substantial gains over the RGB-D baseline under challenging motion, while preserving its performance when event evidence is sparse and remaining competitive with reconstruction-based event tracking.

15 June 2026 in Engineering Research Express
Verified
Journal Article
Closed access
Long ChenLong Chen
Long Chen
CWChenchen Wang
Chenchen Wang
BBBaicheng Bian
Baicheng Bian
Steel strip surface defects generally exhibit blurred boundaries, easily confused category features, and large-scale variations, which severely limit the detection accuracy and robustness of existing methods. To address these issues, this paper regards steel strip surface defects as salient objects from the perspective of salient object detection and proposes a lightweight Two-Stage Category-Aware Refinement Network (TSCARNet) for steel strip surface defect detection. The proposed network consists of an initial defect detection stage and a category-aware refinement stage. In the first stage, the backbone network is redesigned in a lightweight manner, and multi-scale feature interaction in the encoder-decoder process is enhanced to improve the initial localization capability for low-contrast and blurred-boundary defects, thereby generating reliable coarse defect masks. In the second stage, the Category-Aware Separation Enhancement module (CASE) is used to separate and enhance category-specific features, thereby improving the discrimination of easily confused defects. In addition, the Multi-Scale Refinement Module (MSRM) is employed to adaptively capture defect features at different scales under the guidance of coarse masks, thereby enhancing the detection capability for defects with significant scale variations. Experimental results show that, compared with 18 state-of-the-art methods, the proposed network achieves superior performance in both detection accuracy and computational efficiency. On the SD-saliency-900 dataset, the proposed model reduces the Mean Absolute Error (MAE) to 0.0109, representing a 7.6% reduction compared with the best competing method. Meanwhile, it achieves a running speed of 351 frames per second (FPS) on a single GPU, demonstrating a favorable balance between detection accuracy and real-time performance.