---

name: media-processing description: "Use MMCV image, video, optical-flow, visualization, and array quantization utilities safely without reopening MMCV source docs." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MMCV Media Processing

Use this sub-skill when a task asks for MMCV media utilities: reading or writing images, decoding image bytes, resizing, cropping, flipping, padding, rotating, normalizing or denormalizing arrays, color conversions, photometric helpers, converting tensors to images, reading or writing video/frames, ffmpeg-backed video edits, dense or sparse optical flow IO/quantization/warping, drawing boxes, converting flow to RGB, or scalar array quantization.

Read First

• references/api-reference.md for compact signatures, accepted parameters, return shapes, and the most common gotchas.
• references/workflows.md for copyable recipes that combine image, video, flow, visualization, and quantization calls.
• references/troubleshooting.md for symptoms and recovery steps around color order, shape order, headless display, codecs, optional backends, dtype/range, and flow shape issues.
• scripts/media_smoke_check.py to verify the installed mmcv media surface with deterministic tiny arrays and temporary files.

Scope Boundaries

• Use this sub-skill for functional APIs in mmcv.image, mmcv.video, mmcv.visualization, and mmcv.arraymisc.
• Route transform pipeline classes such as Resize, Compose, and LoadImageFromFile to ../data-transforms/.
• Route CNN/layer builders to ../cnn-model-building/.
• Route compiled mmcv.ops availability, build failures, and _ext import issues to ../ops-and-builds/; this media surface works in mmcv-lite when its Python dependencies are present.

Safety Defaults

• Prefer show=False plus out_file=... or returned arrays for visualization in headless sessions; avoid imshow() and flowshow() unless a GUI display is definitely available.
• Treat image arrays as H, W, C, but many MMCV size parameters are (width, height) and padding target shape is (height, width).
• Treat color tuples and default image reads as BGR unless you explicitly request channel_order='rgb' or call a conversion helper.
• Use backend_args rather than deprecated file_client_args when reading from non-local file backends.

Quick Validation

Run the bundled check in any environment where MMCV is installed:

python scripts/media_smoke_check.py --help
python scripts/media_smoke_check.py

The script checks imports, image bytes round-trip, BGR/RGB conversion, resizing scale reports, normalization/denormalization, array quantization, and optional flow helpers without reading the original repository checkout.