视觉故障与回归检测
Visual glitch & regression detection
Spot clipping, rendering and HUD regressions.
20 条本项目编写的提案 · 分类 v1.1 · 新增基准运行 0 次。下方验证方式与部分得分规则是设计，不是执行结果。
- 典型应用
- game captures · engine scenes
- 小类 GUI 需求标签
G2· GUI 必要
任务提案 · 20
展开查看完整原文与建议检查方式；两种界面均保留提案英文原文。
R13.03.01.02-T01 Flag scenes with missing objects between two builds' screenshots
原始指令
input/captures/ holds 10 pairs of screenshots of the same camera shots from build N−1 and build N of a Unity game. Some build-N shots are missing objects (props, NPCs, effects). In the self-hosted review tool, mark each pair as 'regression' or 'ok' and draw a box around each missing object on the build-N image.
- 平台
web· 任意系统（网页）- 应用
- game captures · review tool
- 交互表面
brow· 浏览器- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
vis· 视觉判断drag· 拖放click· 指向与点击- 工作产物
WP10· 标注- 条件
- 未声明
- 提案难度
- 3 / 5
建议验证方式
Review-tool DB: per-pair verdict F1 vs ground truth; boxes matched to ground-truth missing-object regions at IoU ≥ 0.5. Dynamic elements (clouds, particles) differ in every pair and are not regressions. Terminal and Python are blocked on the VM and the review tool issues no API token, so pairs must be compared by eye in the browser.
建议部分得分
50% verdict F1; 50% box recall at IoU ≥ 0.5
R13.03.01.02-T02 Timestamp a wall-clipping glitch in a gameplay video
原始指令
Open input/run_042.mp4 (60 s gameplay capture, 60 fps) in VLC or the review tool. The character clips through a wall once. Report the start time of the clip-through in answer.json to the nearest 0.5 s and the frame where it is most visible as a PNG.
- 平台
mac· macOSwin· Windows- 应用
- game captures · VLC
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
av· 音视频时序vis· 视觉判断nav· 平移 / 缩放 / 滚动- 工作产物
WP12· 答案 / 决策- 条件
- 未声明
- 提案难度
- 3 / 5
建议验证方式
answer.json time within ±2 s of ground truth (full credit within ±0.5 s); PNG must perceptually match a frame inside the labelled glitch interval (pHash distance ≤ 6). Terminal, Python and ffmpeg are blocked on the VM, so the glitch must be found and the frame exported by scrubbing in VLC or the review tool.
建议部分得分
60% scaled by timestamp error (2 s → 0.5 s); 40% correct frame exported
R13.03.01.02-T03 Check HUD elements against a visual spec
原始指令
input/hud_spec.md says: health bar top-left, red (#D32F2F ±10 ΔE), ammo counter bottom-right showing '30', minimap top-right 180 px square. Check the 6 screenshots in input/hud/ against the spec and fill hud_check.csv (screenshot, element, pass/fail).
- 平台
mac· macOSwin· Windows- 应用
- game captures
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
vis· 视觉判断type· 文字与表单- 工作产物
WP02· 记录 / 表格- 条件
- 未声明
- 提案难度
- 2 / 5
建议验证方式
hud_check.csv compared to ground-truth verdicts (18 element checks). Seeded failures: ammo shows '3O', minimap 160 px, bar colour shifted. Terminal and Python are blocked on the VM, so elements must be checked by viewing the images with an image viewer and its colour picker.
建议部分得分
50% scaled by correct verdicts on screenshots 1-3 (9 checks); 50% scaled by correct verdicts on screenshots 4-6 (9 checks)
R13.03.01.02-T04 Frame the camera on a reported Z-fighting artifact in Unity
原始指令
Open the Unity project fixtures/CityBlock in the Unity Editor. Bug report: 'Z-fighting on the north facade sign of building B3 when viewed from the plaza at distance ~25 m.' Move the Scene/Game camera to reproduce it so the flicker is visible, save a screenshot to Captures/zfight.png and save the camera transform in the scene.
- 平台
win· Windowsmac· macOS- 应用
- Unity
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
geom· 精确几何nav· 平移 / 缩放 / 滚动vis· 视觉判断- 工作产物
WP10· 标注- 条件
- 未声明
- 提案难度
- 4 / 5
建议验证方式
Saved scene: camera position within the plaza region, 20-30 m from the sign, sign inside the view frustum. Screenshot analysed: the sign region shows interleaved stripe pattern (Z-fight detector on two coplanar materials). Fallback VLM check is flagged as such. The agent has no shell, and a scene file written by any process other than the Unity Editor is detected.
建议部分得分
40% camera in correct region and distance; 30% sign in view; 30% artifact visible in screenshot
R13.03.01.02-T05 List frames with missing-texture placeholders
原始指令
input/captures_tex/ holds 30 screenshots from a test pass. Some show the engine's missing-texture placeholder (magenta/black checker) on one or more objects. Write missing_tex.txt listing the filenames that contain it.
- 平台
mac· macOSwin· Windowslin· Linux- 应用
- game captures
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
vis· 视觉判断- 工作产物
WP12· 答案 / 决策- 条件
- 未声明
- 提案难度
- 2 / 5
建议验证方式
Filename set F1 vs ground truth. Seeded hard negatives include a magenta neon sign and a checkered floor. Terminal and Python are blocked on the VM, so screenshots must be inspected in an image viewer rather than colour-matched by script.
建议部分得分
50% scaled by F1 from 0 to 0.5; 50% scaled by F1 from 0.5 to 1.0
R13.03.01.02-T06 Timestamp LOD popping events in a fly-through capture
原始指令
input/flythrough.mp4 (90 s, 60 fps) is a camera fly-through of an open-world scene. Objects visibly 'pop' when switching level of detail. In the review tool, add a marker at each pop (time and a box on the object).
- 平台
web· 任意系统（网页）- 应用
- game captures · review tool
- 交互表面
brow· 浏览器- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
av· 音视频时序vis· 视觉判断drag· 拖放- 工作产物
WP10· 标注- 条件
- 未声明
- 提案难度
- 4 / 5
建议验证方式
Markers matched to ground-truth pops within ±0.5 s and box IoU ≥ 0.3: recall and precision. Pops were labelled from engine LOD-switch logs. Terminal, Python and ffmpeg are blocked on the VM and the review tool issues no API token, so pops must be marked while watching in the browser.
建议部分得分
60% recall; 25% precision; 15% box quality
R13.03.01.02-T07 Tune shadow bias to remove acne without peter-panning in Godot
原始指令
Open fixtures/shadow_scene in the Godot editor. The directional light causes shadow acne on the terrain, and raising bias too much detaches shadows from the crates (peter-panning). Adjust the light's shadow bias and normal bias so both artifacts are gone, then save the scene and render Captures/shadow.png from the saved camera.
- 平台
win· Windowsmac· macOSlin· Linux- 应用
- Godot
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
geom· 精确几何vis· 视觉判断click· 指向与点击- 工作产物
WP08· 应用 / 系统状态- 条件
- 未声明
- 提案难度
- 4 / 5
建议验证方式
Headless re-render from the saved scene: acne metric (high-frequency shadow noise on lit terrain mask) below threshold and contact-gap metric (unshadowed pixels at crate bases) ≤ 2 px. Edits to .tscn outside the editor detected by editor's scene UID log.
建议部分得分
40% acne removed; 40% no peter-panning; 20% capture saved
R13.03.01.02-T08 Report UI elements clipped at two resolutions
原始指令
Run the Godot desktop build fixtures/menu_build at 1280×720 and 3840×2160 (windowed, set via its Options menu). Visit all 6 menus and report in clipping.csv each UI element that is cut off or overflows the screen at each resolution.
- 平台
win· Windowsmac· macOS- 应用
- engine scenes
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
click· 指向与点击vis· 视觉判断nav· 平移 / 缩放 / 滚动- 工作产物
WP02· 记录 / 表格- 条件
- 未声明
- 提案难度
- 3 / 5
建议验证方式
clipping.csv (menu, element, resolution) F1 vs ground truth (7 seeded issues). Resolution changes must appear in the build's settings log. Terminal and Python are blocked on the VM, so the settings file cannot be edited outside the Options menu and clipping must be judged on screen.
建议部分得分
20% both resolutions exercised; 80% scaled by F1
R13.03.01.02-T09 Find T-pose frames in a cutscene capture
原始指令
input/cutscene.mp4 (2 min, 30 fps) is a cutscene where animation fails at times and characters snap to a T-pose. Report all frame ranges with a T-posed character in tpose.csv (start_frame, end_frame, character).
- 平台
mac· macOSwin· Windows- 应用
- game captures
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
av· 音视频时序vis· 视觉判断- 工作产物
WP10· 标注- 条件
- 未声明
- 提案难度
- 3 / 5
建议验证方式
Range overlap with ground truth: temporal IoU per range ≥ 0.5 counts as a match; character name must match. Recall and precision. Terminal, Python and ffmpeg are blocked on the VM, so frames must be inspected in the player rather than analysed by script.
建议部分得分
60% recall; 25% precision; 15% character names
R13.03.01.02-T10 Detect missing particle effects between two builds
原始指令
input/fx_pairs/ holds 12 short clips (3 s) of the same scripted moments in build N−1 and N (explosions, spells, footsteps dust). Mark in fx_check.csv which build-N clips are missing an effect present in N−1, naming the effect.
- 平台
mac· macOSwin· Windows- 应用
- game captures
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
av· 音视频时序vis· 视觉判断- 工作产物
WP02· 记录 / 表格- 条件
- 未声明
- 提案难度
- 3 / 5
建议验证方式
fx_check.csv verdict F1 vs ground truth; effect name matched against a synonym list. Particle randomness differs between builds even when correct. Terminal, Python and ffmpeg are blocked on the VM, so clips must be compared in the player rather than analysed by script.
建议部分得分
70% verdict F1; 30% effect names
R13.03.01.02-T11 Find a material with an inverted normal map in Unreal
原始指令
Open the Unreal Editor project fixtures/Courtyard. Lighting on one wall material looks wrong (bumps appear as dents). Use buffer visualisation (World Normal) and the material editor to find the material whose normal map has an inverted green channel, fix it by setting 'Flip Green Channel' on the texture, and save.
- 平台
win· Windowsmac· macOS- 应用
- Unreal Engine
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
vis· 视觉判断click· 指向与点击nav· 平移 / 缩放 / 滚动- 工作产物
WP08· 应用 / 系统状态- 条件
- 未声明
- 提案难度
- 5 / 5
建议验证方式
Saved .uasset metadata parsed (via a commandlet in the harness): only the seeded texture has bFlipGreenChannel = true; no other assets modified (package hash diff). Python editor scripting plugin is disabled.
建议部分得分
40% correct material identified (texture changed); 40% fix applied; 20% no other assets changed
R13.03.01.02-T12 Measure a tonemapping colour regression on reference patches
原始指令
input/tonemap/ has captures of an in-game colour chart from build N−1 and N (SDR). For each of the 24 patches, measure mean colour in a patch-centre region using an image editor's colour picker or info panel and list patches with ΔE2000 > 3 between builds in tonemap.csv.
- 平台
mac· macOSwin· Windows- 应用
- game captures · Photoshop · GIMP
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
CL· CLI 足够- 交互原语
vis· 视觉判断geom· 精确几何type· 文字与表单- 工作产物
WP02· 记录 / 表格- 条件
- 未声明
- 提案难度
- 4 / 5
建议验证方式
tonemap.csv patch set vs ground truth computed from the images (patches with ΔE2000 > 3); reported ΔE values within ±0.5. Terminal and Python are blocked on the VM, so sampling must be done in the editor; accuracy of manual sampling is what is measured.
建议部分得分
60% patch set F1; 40% ΔE values within tolerance
R13.03.01.02-T13 Locate flickering geometry in a long capture
原始指令
input/walk_2min.mp4 (2 min, 60 fps) contains intermittent single-frame flicker of a distant building (visibility toggling). Report every flicker event's frame number in flicker.csv.
- 平台
mac· macOSwin· Windows- 应用
- game captures · VLC
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
av· 音视频时序vis· 视觉判断nav· 平移 / 缩放 / 滚动- 工作产物
WP10· 标注- 条件
long· 长流程- 提案难度
- 4 / 5
建议验证方式
Frame numbers matched to ground truth within ±2 frames; recall and precision. Events last 1-2 frames, so frame stepping is required. Terminal, Python and ffmpeg are blocked on the VM, so frames must be stepped through in VLC rather than analysed by script.
建议部分得分
70% recall; 30% precision
R13.03.01.02-T14 Flag overlapping text in localized menu screenshots
原始指令
input/loc_shots/ has 40 screenshots of the same 10 menus in 4 languages (EN, DE, JA, AR). Flag every screenshot where text overlaps other text or UI elements, or runs outside its button, in loc_flags.csv (file, element).
- 平台
mac· macOSwin· Windows- 应用
- game captures
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
vis· 视觉判断- 工作产物
WP10· 标注- 条件
- 未声明
- 提案难度
- 2 / 5
建议验证方式
loc_flags.csv file+element matched to ground truth (9 seeded issues); F1. Right-to-left layout in AR is correct by design and is a hard negative. Terminal and Python are blocked on the VM, so screenshots must be inspected in an image viewer rather than analysed by OCR scripts.
建议部分得分
50% scaled by F1 from 0 to 0.5; 50% scaled by F1 from 0.5 to 1.0
R13.03.01.02-T15 Reproduce a terrain-clipping report in Godot and file it with the camera transform
原始指令
Bug report: 'Camera clips into the hill near the windmill when orbiting at max zoom.' Open fixtures/valley in the Godot editor, run the scene, reproduce the clipping with the in-game orbit camera, capture Captures/clip.png and file a MantisBT note with the camera position and rotation shown in the debug overlay.
- 平台
win· Windowsmac· macOSlin· Linux- 应用
- Godot · MantisBT
- 交互表面
desk· 桌面 GUIbrow· 浏览器- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
geom· 精确几何nav· 平移 / 缩放 / 滚动vis· 视觉判断- 工作产物
WP10· 标注- 条件
multiapp· 多应用- 提案难度
- 4 / 5
建议验证方式
Reported camera position inside the terrain mesh (ray test in harness) near the windmill (≤ 30 m); screenshot shows terrain interior (backface ratio metric). Note added to the issue. Terminal and Python are blocked on the VM, and the harness logs debug-overlay transforms per frame, so the reported transform must be one the run actually reached.
建议部分得分
40% clipping reproduced; 30% transform reported correctly; 30% screenshot and note filed
R13.03.01.02-T16 Identify stutter from duplicated frames in a capture
原始指令
input/pan_60fps.mp4 is a 30 s capture of a smooth camera pan recorded at 60 fps. Frame pacing problems show up as duplicated or skipped frames. Report the time ranges with stutter in stutter.csv (start_s, end_s).
- 平台
mac· macOSwin· Windows- 应用
- game captures · VLC
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
CL· CLI 足够- 交互原语
av· 音视频时序vis· 视觉判断- 工作产物
WP10· 标注- 条件
- 未声明
- 提案难度
- 4 / 5
建议验证方式
Ranges matched to ground truth (derived from engine frame-time log) with temporal IoU ≥ 0.5; recall and precision. Terminal, Python and ffmpeg are blocked on the VM, so duplicated frames must be found by stepping in the player rather than by frame hashing.
建议部分得分
70% recall; 30% precision
R13.03.01.02-T17 Approve or reject 25 visual-regression diffs
原始指令
A self-hosted visual-regression dashboard (reg-suit style) shows 25 diffs between golden images and the new build, each with a changelog note. Approve diffs that match an intended change in the changelog and reject the rest as regressions with a one-line reason.
- 平台
web· 任意系统（网页）- 应用
- visual regression dashboard
- 交互表面
brow· 浏览器- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
vis· 视觉判断click· 指向与点击type· 文字与表单- 工作产物
WP12· 答案 / 决策- 条件
- 未声明
- 提案难度
- 3 / 5
建议验证方式
Dashboard DB: approve/reject per diff vs ground truth; accuracy. Reasons are recorded; rejection reasons spot-checked by rubric. The dashboard issues no API token and accepts verdicts only from the CSRF-bound browser session; Terminal and Python are blocked, so diffs cannot be scored by script.
建议部分得分
85% scaled by verdict accuracy; 15% reasons present
R13.03.01.02-T18 Check HUD safe-area violations on notched device profiles
原始指令
On the Android emulator, install fixtures/hud_test.apk and run it on 3 device profiles (Pixel 7, a 20:9 notch device, a tablet). Screenshot the HUD in each and report in safe_area.csv any element overlapping the cutout or rounded corners.
- 平台
and· Android- 应用
- engine scenes
- 交互表面
mob· 移动触控- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
touch· 触控手势vis· 视觉判断- 工作产物
WP02· 记录 / 表格- 条件
- 未声明
- 提案难度
- 3 / 5
建议验证方式
safe_area.csv (device, element) F1 vs ground truth; screenshots for each profile must exist with the right resolution. Python and image-analysis tools are not installed on the agent VM, so overlaps with the cutout and corners must be judged by eye.
建议部分得分
30% three profiles run with screenshots; 70% F1
R13.03.01.02-T19 Use the Unity Frame Debugger to fix an object that disappears
原始指令
In Unity project fixtures/Shop, the glass display case vanishes when viewed in front of the skybox. Use the Frame Debugger to find the draw call ordering problem, fix the material's render queue or transparency settings in the Inspector so it renders correctly, and save.
- 平台
win· Windowsmac· macOS- 应用
- Unity
- 交互表面
desk· 桌面 GUI- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
vis· 视觉判断click· 指向与点击nav· 平移 / 缩放 / 滚动- 工作产物
WP08· 应用 / 系统状态- 条件
- 未声明
- 提案难度
- 5 / 5
建议验证方式
Headless player render from the saved project: case visible (pixel mask overlap ≥ 0.9 vs reference). Material YAML diff limited to the glass material's queue/blend fields; other assets unchanged. The agent has no shell, and a material file written by any process other than the Unity Editor is detected.
建议部分得分
30% correct material identified; 50% case renders correctly; 20% no unrelated asset changes
R13.03.01.02-T20 Triage 120 screenshot pairs for regressions with noisy dynamic content
原始指令
A review tool lists 120 screenshot pairs (golden vs nightly) across 12 levels; weather, crowds and time of day vary between captures. Mark each pair 'regression' or 'ok' and, for regressions, box the region and pick a category (missing, wrong material, lighting, UI, geometry).
- 平台
web· 任意系统（网页）- 应用
- game captures · review tool
- 交互表面
brow· 浏览器- GUI 需求 · 任务设定
G2· GUI 必要- GUI 需求 · 普通权限
G2· GUI 必要- 交互原语
vis· 视觉判断drag· 拖放click· 指向与点击- 工作产物
WP10· 标注- 条件
long· 长流程- 提案难度
- 5 / 5
建议验证方式
Verdict precision/recall (target ≥ 0.9 each); boxes IoU ≥ 0.5; category accuracy. 21 seeded regressions; dynamic noise is the hard negative. Terminal and Python are blocked on the VM and the review tool issues no API token, so pairs must be judged by eye in the browser.
建议部分得分
50% verdict F1; 30% box recall; 20% category accuracy
已有研究映射
以下沿用研究目录中的基准与小类映射，不是逐题标注、执行记录或产品背书。