DeepstreamDemo

Incremental component playground for proving out a DeepStream-based replacement for the production GStreamer + nvcodec + Qt OpenGL pipeline.

Philosophy

This repo deliberately avoids being one big application. Each subdirectory under components/ is an isolated, independently buildable, independently runnable prototype that proves exactly one capability. A component only graduates to "ready for integration" after its success criteria (in its own README.md) are met on real hardware.

Integration is the disease this repo tries to cure. We will not write a multi-thousand-line app and then debug it as a unit. We will land each piece on its own, prove it, document its contract, and only then compose.

Layout

DeepstreamDemo/
├── README.md                       this file
├── docs/
│   ├── architecture.md             the sub-project rules
│   └── component-template.md       what every new component must contain
└── components/
    ├── 01_qt_eglimage_zerocopy/    P0.1: NVMM → Qt OpenGL GPU-only display
    ├── 02_nvurisrcbin_reconnect/   P0.2: single RTSP source + reconnect
    ├── 03_latency_probe/           P0.3: source→paint latency measurement contract
    ├── 04_multi_rtsp_widgets/      P0.4: independent multi-RTSP widget proof
    ├── 05_selective_yolo_infer/    P0.5: per-stream selective YOLO inference
    ├── 06_dual_4k_stitch_restream/ P0.6: dual 4K GPU stitch + restream
    └── 07_full_pressure_integration/ P0.7: 8x 1080p + 1x 4K YOLO + 2x 4K stitch in one app

Component Index

#	Name	Proves	Status
01	qt_eglimage_zerocopy	NVMM GstBuffer → CUDA-GL interop → Qt QOpenGLWidget	ready
02	nvurisrcbin_reconnect	Single RTSP source with reliable reconnect over 01	ready
03	latency_probe	Reproducible source→paint latency segments (decode/queue, upload/paint, end-to-end) usable as a baseline by 04+	ready
04	multi_rtsp_widgets	Multiple independent RTSP streams displayed in separate Qt widgets with per-stream reconnect/failure isolation	draft
05	selective_yolo_infer	Per-stream selective YOLO inference (nvstreammux/nvinfer/nvdsosd) on a subset of P0.4 panels with runtime mode flips and failure isolation	ready
06	dual_4k_stitch_restream	Two 4K RTSP inputs stitched top-bottom in GPU device memory, previewed locally, and restreamed via RTSP RECORD	ready
07	full_pressure_integration	First integration component: 8x 1080p raw + 1x 4K YOLO + 2x 4K stitch concurrently in one Qt app	ready

How to add a new component

1. Copy docs/component-template.md into components/NN_descriptive_name/README.md.
2. Fill in the Goal, Architecture, Success Criteria, and Known Gotchas sections.
3. Add a self-contained CMakeLists.txt. Do not depend on a top-level CMake project — each component must build standalone.
4. Land the code. Run it. Tick off the success criteria in the README before declaring it ready.

See docs/architecture.md for the discipline rules.