/deepstream-sop

§ 5b — Custom C++ Postprocess Plugin

Reference files to copy (do NOT regenerate):

• nvds_custom_postprocess_tensor_reference.cpp → nvds_action_detector/custom_postprocess/nvds_custom_postprocess_tensor.cpp
• Makefile_custom_postprocess_reference → nvds_action_detector/custom_postprocess/Makefile

Pipeline stage: Stage 1 post-processing — converts DDM raw output tensor into NvDsObjectMeta entries that InferenceOutputTensorParser (§ 3) can read

Why This Must Be Copied Verbatim

nvinferserver requires libnvds_custom_postprocess_tensor.so, compiled from C++ during Docker build (RUN cd custom_postprocess/ && make). The .cpp and Makefile must exist in the source tree or Docker build fails.

The C++ plugin uses DeepStream's nvdsinferserver::IOptions API which has distinct methods for different data types:

• getObj() for single-pointer types (GstBuffer*, NvBufSurface*, NvDsBatchMeta*)
• getValueArray() for vector types (std::vector<NvDsFrameMeta*>, std::vector<uint64_t>)
• getInt() for integer values (unique_id)

Using the wrong method (e.g. getValueArray() on a GstBuffer*) causes a template deduction failure at compile time.

Dockerfile Requirements (both stages needed)

# Stage 1: Extract DeepStream headers (needed by Makefile -I flags).
# Same ARG-based override as Dockerfile_reference — override via
# --build-arg DS_IMAGE=<your DS 9.1 image> while DS-9.1 is not yet public.
ARG DS_IMAGE="nvcr.io/nvidia/deepstream:9.1-triton-multiarch"
FROM ${DS_IMAGE} AS ds_dev

# Stage 2: main build
FROM ${BASE_IMAGE} AS base
...
# Copy DS headers so Makefile can find infer_custom_process.h, nvdsmeta.h, etc.
COPY --from=ds_dev /opt/nvidia/deepstream/deepstream/sources/includes \
    /opt/nvidia/deepstream/deepstream/sources/includes

# Copy application source (must include custom_postprocess/ with .cpp + Makefile)
RUN --mount=type=bind,source=.,target=/tmp/ds_sop \
    cd /tmp/ds_sop && docker/copy_sources.sh ./ /opt/nvidia/nvds_sop/

# Compile the plugin — MUST come after source copy
RUN cd /opt/nvidia/nvds_sop/nvds_action_detector/custom_postprocess/ && make

C++ API Pattern (CRITICAL — build will fail if wrong)

DO NOT include gstnvdsmeta.h — it transitively pulls <gst/gst.h> which is unavailable under the Makefile's include paths, causing a fatal compilation error. Use only DeepStream inference headers + forward-declare GstBuffer:

#include "infer_custom_process.h"   // IInferCustomProcessor, IBatchArray, IOptions
#include "nvbufsurface.h"           // NvBufSurface, NvBufSurfaceParams
#include "nvdsmeta.h"               // NvDsBatchMeta, NvDsFrameMeta, NvDsObjectMeta

typedef struct _GstBuffer GstBuffer;   // forward-declare — do NOT #include <gst/gst.h>

IOptions API Method Selection

This is the #1 compilation failure cause:

Using getValueArray() on a single-pointer type causes: template argument deduction/substitution failed: mismatched types 'std::vector<_Tp>' and 'GstBuffer*'

What the Plugin Does

nvds_custom_postprocess_tensor.cpp implements IInferCustomProcessor:

DDM output tensor (float scores)
        │
        ▼
inferenceDone() — for each score in the batch:
  ├── Creates NvDsObjectMeta:
  │     object_id = frame_num - OFFSET + i
  │         (OFFSET = FRAMES_PER_SIDE + SEQUENCE_BATCH - 1)
  │     confidence = boundary score (float)
  │     class_id = 0 ("boundary")
  │
  ├── Attaches object metas to NvDsFrameMeta
  │     → InferenceOutputTensorParser (§ 3) reads via frame_meta.object_items
  │
  ├── Uses hasValue() guard before each getObj()/getValueArray() call
  │
  └── Acquires/releases batchMeta lock around nvds_add_obj_meta_to_frame()