Second-Year Undergraduate / Shandong Normal University / Computer Science

Xinchun Wang 王新春 · Building efficient and reliable vision systems from undergraduate research.

I am a second-year undergraduate in Computer Science and Technology at Shandong Normal University, working across computer vision, COD, remote sensing super-resolution, and reliable lightweight perception. My current record includes three first-author manuscripts across active review and submission, one published IEEE TGRS Vol. 64 (2026) paper as third author, and an active Huawei Spark Award enterprise challenge on 3D SBS video synthesis.

KBS SCI Q1 First author Under review IEEE GRSL · DABit-Mamba First author Submission in progress IEEE TGRS · Vol. 64 Third author Published

View Achievements Read Research

Academic Snapshot Stage 01 / Reliable Vision

02 Identity Second-year CS undergraduate Computer Science & Technology

SDNU Institution Shandong Normal University 山东师范大学

1st First-author work

KBS SCI Q1 / under review IEEE GRSL · DABit-Mamba First author / submission in progress LLDG First author / submission in progress

Three first-author manuscripts spanning reliable vision, remote sensing, and uncertainty modeling.

TGRS Published achievement FS-FlowNet is published in IEEE TGRS Vol. 64 (2026), with Xinchun Wang listed as third author. Published / Vol. 64 / Third author

Now tackling

3D Huawei enterprise challenge for the Spark Award: 3D SBS video synthesis. Engineering-focused visual generation / side-by-side stereo video / enterprise problem solving

Research Thesis

My work asks one question: how can vision systems stay reliable when the visual signal is weak?

This page is the bridge between my identity and my papers. It explains the common logic behind COD, remote sensing super-resolution, TGRS publication work, and the Huawei Spark Award challenge: make visual systems robust, efficient, and useful under real constraints.

Core Position Reliable perception is not just about stronger models. It is about knowing when, where, and how much a vision system should adjust.

Problem: Weak Signals

Many visual tasks fail when useful cues are faint: camouflaged objects blend into the background, remote sensing images lose fine structure, and generated stereo videos must preserve geometric consistency.

COD RSISR 3D Video

Principle: Controlled Adjustment

My first-author record includes CTTA under review, DABit-Mamba targeting IEEE GRSL, and LLDG in submission; each manuscript discloses its first page, authorship, affiliation, and current status.

CTTA DABit-Mamba LLDG

Evidence: Publication Track

The research direction is supported by concrete progress: three first-author manuscripts across active review and submission, plus a published IEEE TGRS paper as third author.

First Author IEEE TGRS Published

Frontier: Enterprise 3D

The Huawei Spark Award enterprise problem pushes this thesis into application: 3D SBS video synthesis requires reliable visual generation, stereo consistency, and engineering feasibility.

Huawei Spark Award 3D SBS

Representative First-Author Work

Turning the research thesis into a concrete method, benchmark, and manuscript.

This page explains why CTTA is my representative first-author work: it starts from a real failure mode, proposes a bounded correction mechanism, and verifies the gain under explicit comparison constraints.

KBS SCI Q1 Submission / First Author / COD

Constrained Test-Time Adjustment for Camouflaged Object Detection

CTTA studies how to improve camouflaged object detection under open-world test conditions without updating model parameters. The key idea is to adjust only where correction is needed, while preserving predictions that are already correct.

Problem Open-world COD can improve hard cases but damage already-correct samples.

Method ARC anchors output correction, HBR allocates risk, and DMEP verifies fair attribution.

Evidence Regression on low-difficulty cases drops from 28.8% to 13.3% under frozen parameters.

CTTA Camouflaged Object Detection Test-Time Output Adjustment Open-World Stability

First-page preview All Achievements

Performance Evidence Table 1 + Table 2

mean_ds Fwβ 0.8994 +0.0034 vs baseline

mean_ds MAE 0.0153 lower is better

Low-difficulty regression 28.8% -> 13.3% already-correct samples are better protected

Dataset	Baseline	CTTA	ΔFwβ
CHAMELEON	0.9053 / 0.0121	0.9072 / 0.0118	+0.0019
CAMO	0.8998 / 0.0236	0.9045 / 0.0228	+0.0047
COD10K	0.8716 / 0.0110	0.8757 / 0.0107	+0.0041
NC4K	0.9071 / 0.0160	0.9103 / 0.0158	+0.0032

Matched comparisons

Under matched compute and matched adjustment, CTTA keeps the same 13.3% low-difficulty regression rate while preserving positive high-difficulty gains.

Method Guardrails ARC -> HBR -> DMEP

ARC Anchor reliable outputs locks in confident mask structure before adjustment

HBR Route high-risk cases spends correction budget where uncertainty is useful

DMEP Verify attribution keeps gains tied to the proposed correction path

Frozen parameters Output-space adjustment Regression-aware evaluation

Personal Achievements

Four research records across reliable perception, remote sensing, and uncertainty modeling.

IEEE TGRS · Vol. 64 · 2026 Third author · Published

FS-FlowNet: Frequency–Spatial Dual-Domain Residual Flow Network for Remote Sensing Dehazing

Yucheng Xin; Kangjie Huang; Xinchun Wang; Dianjie Lu; Guijuan Zhang; Ruize Wu; Zhuoran Zheng IEEE Transactions on Geoscience and Remote Sensing · Vol. 64 (2026) · pp. 1–15

This is the formal IEEE publication record. The links below resolve to the DOI and IEEE Xplore document; no arXiv version or locally hosted full-text PDF is implied.

Published Vol. 64 (2026) Third author Document 11390698

DOI 10.1109/TGRS.2026.3663290

Open DOI IEEE Xplore record

KBS SCI Q1 First author

Constrained Test-Time Adjustment for Camouflaged Object Detection

A constrained output-space adjustment framework for camouflaged object detection, designed to improve hard cases while protecting already-correct predictions under frozen model parameters.

CTTA ARC HBR DMEP

First-author manuscript under review CTTA for COD / 28.8% -> 13.3%

First page only

IEEE GRSL · Manuscript First author · Submission in progress

DABit-Mamba: Degradation-Aware Bit-Scalable Mamba for Lightweight Remote Sensing Image Super-Resolution

Xinchun Wang; Feng Gao; Yucheng Xin* Xinchun Wang & Yucheng Xin — Shandong Normal University · Feng Gao — Ocean University of China · * Corresponding author

The target venue is IEEE Geoscience and Remote Sensing Letters (GRSL). The first page, authorship, affiliation, and submission status are disclosed here; the remaining manuscript and source stay private.

First author Submission in progress First page public Full text private

Public record First-page preview published

First-page preview

Manuscript First author · Submission in progress

Latent label distribution grid representation for modeling uncertainty

Xinchun Wang; Yucheng Xin* Shandong Normal University · * Corresponding author

The first page is disclosed as an internal-review draft together with verified authorship, affiliation, and submission status. The complete PDF, source, remaining pages, and target venue remain private.

First author Submission in progress First page public Full text private

Public record Internal-review first page

First-page preview

Now Running

Current AAAI experiments extending reliable perception beyond the training set.

AAAI Submission in Progress

From hidden targets to controlled test-time behavior.

This section carries the trajectory of the paper forward: how to make perception systems stay reliable when cues are weak, targets are hard, and naive correction can damage already-correct predictions.

Problem: robust visual understanding under weak cues and distribution shift
Direction: constrained correction, risk-aware allocation, and stable inference behavior
Status: AAAI-oriented model design, ablations, and evaluation protocol tuning in progress

lab/monitor running

benchmark scope 4 sets

parameter update frozen

adjustment policy bounded

next target AAAI

Milestones

A research trajectory shaped by hidden-object perception and reliable test-time reasoning.

Start

Entered CV research through segmentation, perception, and difficult visual scenes with weak structure.

Submissions

Advanced first-author manuscripts for KBS, IEEE GRSL (DABit-Mamba), and LLDG across active review and submission.

Now

Building on the published FS-FlowNet record in IEEE TGRS Vol. 64 (2026), with Xinchun Wang listed as third author.

Contact

Open to collaboration on COD, segmentation, and reliable test-time perception.

Reach me directly by email or GitHub for research collaboration, paper discussion, code, and future projects growing out of CTTA.

1198383781w@gmail.com GitHub Google Scholar