Hi,
We are seeing intermittent vertical line artifacts on camera images in a DriveOS + NvSIPL pipeline. Below is a detailed description of the environment and the symptom. Any similar known issues or guidance would be greatly appreciated.
1. Environment / pipeline
- Platform: NVIDIA DriveOS on DRIVE AGX / Orin
- Camera frontend: Sensor → Deserializer (e.g. MAX967x class) → NVCSI / VI → NvSIPL
- ISP output used by the application: ISP2
- ISP / pipeline native image format: YUV420 NV12 (semi-planar: Y plane + interleaved UV plane), not RGB at the ISP output
- Configured image size in our application: 3848 × 2168
- Multi-camera: on the order of ~4 ISP2 streams per process; two processes may run on the same SoC and share ISP / VI / NVCSI resources (roughly ~8 concurrent 4K-class ISP2 streams in the worst case)
- By default we disable ISP0/ISP1 and only enable ISP2 output
Data path after ISP:
Sensor / SerDes / NVCSI / VI
→ NvSIPL ISP2 (YUV420 NV12, \~3848×2168)
→ S1 application (Producer): full-frame copy of SIPL NvSciBufObj
→ NvSciStream C2C over PCIe
→ S2 application (Consumer): receive YUV NV12
→ YUV(NV12) → BGR conversion
→ JPEG / PNG save (artifacts observed on these saved images)
Important format note:
- The camera pipeline output we consume is YUV (NV12) from the beginning.
- BGR / JPEG exist only for display and storage after color conversion on S2.
- The defect is observed on the saved images after YUV→BGR; we are trying to determine whether it already exists in the ISP2 NV12 buffer before C2C / before color conversion.
2. Symptom overview
- Artifact type: very thin vertical lines superimposed on the image
- Color:
- Most common: green vertical segments
- Sometimes: a black vertical line in the same frame (or in other abnormal frames)
- Frequency: roughly on the order of ~0.5% of frames (order-of-magnitude estimate from field observation); highly intermittent, no reliable forced reproduction steps so far
- Not every abnormal frame has the same number of lines: a frame may show one or multiple green segments; black lines are less frequent than green
3. Geometric / visual characteristics
- The lines are perfectly straight and vertical (aligned with the image column axis)
- They do not follow the fisheye / wide-angle lens geometric distortion of real scene content
→ they behave like digital / pipeline artifacts, not physical objects (cables, poles, etc.) in the scene - Horizontal position:
- Appears roughly stable / periodic
- Visually clustered near ~1/4, 1/2, and 3/4 of the image width
- For an effective width of 3840, that corresponds approximately to:
- x ≈ 960
- x ≈ 1920
- x ≈ 2880
- Adjacent abnormal column positions are separated by about ~960 pixels
- Vertical extent:
- Often a short vertical segment, not always spanning the full image height
- Frequently appears in the middle / lower region of the frame
- Start/end rows can vary between occurrences
- Line thickness: very narrow (appears close to 1-pixel-class / very thin column)
4. Color behavior (especially the green line)
- The green line does not look like a constant pure-green overlay (fixed RGB green painted on top)
- Along the line, brightness tends to follow the local background:
- over bright areas → the green segment looks brighter
- over dark areas → the green segment looks darker
- This appearance is more consistent with a local chroma (U/V) anomaly on NV12, while luma (Y) may still be largely plausible, rather than a hard-coded green RGB draw
The black line, when present:
- Also appears as a thin, perfectly vertical digital column
- Looks closer to a dead / zeroed column (missing or cleared samples) than a chroma-tint line
5. Relationship to our application processing
- Our S1 producer performs a full-frame copy from the SIPL buffer into the NvSciStream packet
- We do not intentionally slice / process the image into width/4 vertical strips in application code
- Despite that, the defect locations still align with quarter-width positions (~960-pixel period)
- Classic row-major CPU memcpy tearing usually produces horizontal bands; this defect is vertical and position-periodic, which makes a simple full-frame row-copy tear a weaker explanation for the quarter-width pattern by itself
6. Resolution / geometry detail that may be relevant
- Application constants use WIDTH = 3848, HEIGHT = 2168
- The recurring X positions match 3840 / 4 = 960 better than 3848 / 4 = 962
- Note that 3848 ≈ 960×4 + 8, i.e. there are 8 extra pixels beyond 3840
- We therefore suspect a possible mismatch among:
- active / effective width (e.g. 3840)
- configured / reported width (3848)
PlaneWidth/PlanePitch/ ROI / crop / padding
but this is still an observation from the symptom pattern, not a proven root cause
7. Runtime / system context when observed
- Observed on field / station monitoring images (ceiling-mounted wide-angle / fisheye-style views of an industrial platform)
- Can appear while people / equipment are present in the scene; the lines cut across people, floor markings, and mechanical structures without respecting scene perspective
- The issue can occur during otherwise “normal” streaming (not only during obvious camera bring-up / teardown)
- We have separately seen ISP / ACP style failures on this platform (e.g. SIPL
NOTIF_ERROR_ACP_PROCESSING_FAILURE), but we do not currently have a proven one-to-one correlation that every vertical-line frame is accompanied by an ACP/ISP failure notification - In other words: this can look like a silent image-quality defect, not always a hard pipeline error event
8. What the problem is / is not (based on observation only)
Consistent with observations:
- Intermittent digital vertical artifacts on YUV-originated images
- Strong spatial preference for quarter-width columns
- Green line behaves like chroma-related corruption; black line like a cleared/dead column
Less consistent with a simple explanation from the visual evidence alone:
- A real physical line in the scene (geometry / fisheye mismatch)
- A permanent sensor dead column (because it is intermittent and length varies)
- A UI / debug overlay drawn by the application at fixed HUD positions
9. Ask to the community / NVIDIA
- Has anyone seen similar intermittent vertical lines at regular ~width/4 boundaries on NvSIPL ISP2 NV12 output under DriveOS?
- For NV12, is a thin green vertical segment with background-dependent brightness a typical signature of corrupted UV at a fixed chroma column?
- Are there known issues or constraints around width vs pitch (e.g. 3848 vs 3840) that could produce periodic vertical chroma artifacts?
- Which DriveOS / NvSIPL / kernel logs are most relevant when the defect is mostly silent (no obvious ISP/ACP notification on every occurrence)?
We can share cropped sample images and more exact pixel coordinates / version strings if needed.
Thanks in advance for any help.
