Add openvino_hggd module: HGGD OpenVINO iGPU enablement

.github/labeler.yml

@@ -31,6 +31,10 @@
 - changed-files:
   - any-glob-to-any-file: 'modules/openvino_flashocc/**/*'
 
+'category: hggd':
+- changed-files:
+  - any-glob-to-any-file: 'modules/openvino_hggd/**/*'
+
 'category: custom operations':
 - changed-files:
   - any-glob-to-any-file: 'modules/custom_operations/**/*'

.github/labels.yml

@@ -28,6 +28,10 @@
   color: "#d3d3d3"
   description: "OpenVINO FlashOCC module (modules/openvino_flashocc)"
 
+- name: "category: hggd"
+  color: "#d3d3d3"
+  description: "OpenVINO HGGD module (modules/openvino_hggd)"
+
 - name: "category: build"
   color: "#d3d3d3"
   description: "OpenVINO cmake script / infra"

README.md

@@ -22,6 +22,7 @@ This list gives an overview of all modules available inside the contrib reposito
 * [**Ollama-OpenVINO**](./modules/ollama_openvino): OpenVINO GenAI empowered Ollama which accelerate LLM on Intel platforms(including CPU, iGPU/dGPU, NPU).
 * [**ov_training_kit**](./modules/ov_training_kit): Training Kit Python library -- provides scikit-learn, PyTorch and Tensorflow wrappers for training, optimization, and deployment with OpenVINO on AI PCs.
 * [**OpenVino BEVFusion**](./modules/openvino_bevfusion): OpenVino supported implementation of the BEVFusion model.
+* [**OpenVINO HGGD**](./modules/openvino_hggd): OpenVINO-accelerated HGGD grasp detection using custom GPU point cloud extensions.
 * [**OpenVINO FlashOCC**](./modules/openvino_flashocc): OpenVINO-supported FlashOCC export, inference, and benchmarking pipeline for 3D occupancy prediction.
 * [**OpenVINO Notes**](./modules/openvino-notes): Modular Android notes application foundation with local storage, Compose UI, and adapter contracts for Google identity, Drive sync, and OpenVINO assistance.
 * [**3D**](./modules/3d): A collection of 3D vision models implemented with OpenVINO. Currently, it includes the following models:
@@ -53,6 +54,7 @@ Additional build instructions are available for the following modules:
 * [**openvino-langchain**](./modules/openvino-langchain): LangChain.js integrations for OpenVINO™
 * [**OpenVINO Notes**](./modules/openvino-notes/README.md#build-and-test): Standalone Android Gradle build, prerequisites, architecture checks, and unit tests.
 * [**OpenVino BEVFusion**](./modules/openvino_bevfusion): Check the [INSTRUCTIONS](./modules/openvino_bevfusion/EXPORT_AND_INFERENCE_GUIDE.md) for detailed usage and build instructions.
+* [**OpenVino HGGD**](./modules/openvino_hggd): Check the [INSTRUCTIONS](./modules/openvino_hggd/README.md) for detailed usage and build instructions.
 * [**OpenVINO FlashOCC**](./modules/openvino_flashocc): Check the [INSTRUCTIONS](./modules/openvino_flashocc/EXPORT_AND_INFERENCE_GUIDE.md) for detailed usage and build instructions.
 * **3D**:
   * [**Point Pillars**](./modules/3d/pointPillars): Check the [README](./modules/3d/pointPillars/README.md) for detailed usage and build instructions.

modules/openvino_hggd/.gitignore

@@ -0,0 +1,34 @@
+# Copyright (C) 2018-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+# OpenVINO exported models
+openvino_models/
+
+# Inference outputs
+output/
+*.npy
+
+# Logs
+*.log
+
+# Compiled extensions
+*.so
+
+# Generated visualizations
+viz_*.png
+
+# Python
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+
+# Virtual environments
+.venv/
+venv/
+env/
+
+# Build artifacts
+build/
+dist/
+*.egg-info/

modules/openvino_hggd/README.md

@@ -0,0 +1,103 @@
+<!--
+Copyright (C) 2018-2026 Intel Corporation
+SPDX-License-Identifier: Apache-2.0
+-->
+
+# HGGD — OpenVINO iGPU Enablement
+
+OpenVINO inference enablement for [HGGD](https://github.com/THU-VCLab/HGGD) (Hybrid Grasp Detection and Generation) on Intel integrated GPU.
+
+## Prerequisites: Fetch Upstream HGGD Files
+
+The `customgraspnetAPI`, `dataset`, and `models` directories are not included in this module — they are identical to the upstream repository and should be copied directly from there:
+
+```bash
+git clone https://github.com/THU-VCLab/HGGD /tmp/HGGD_upstream
+cp -r /tmp/HGGD_upstream/customgraspnetAPI path_to_openvino_hggd/
+cp -r /tmp/HGGD_upstream/dataset          path_to_openvino_hggd/
+cp -r /tmp/HGGD_upstream/models           path_to_openvino_hggd/
+```
+
+## Setup
+
+```bash
+cd path_to_openvino_hggd
+bash setup.sh
+conda activate hggd_intel
+```
+
+## Export Models
+
+```bash
+cd path_to_openvino_hggd
+conda activate hggd_intel
+
+python export_models.py \
+  --checkpoint-path /path/to/HGGD_realsense_checkpoint \
+  --output-dir openvino_models \
+  --ov-device GPU
+```
+
+## Run
+
+```bash
+cd path_to_openvino_hggd
+conda activate hggd_intel
+
+bash run.sh \
+  /path/to/HGGD_realsense_checkpoint \
+  /path/to/dataset/6dto2drefine_realsense \
+  /path/to/graspnet \
+  100 GPU
+```
+
+## Evaluation
+
+Inference dumps per-frame grasp predictions under
+`output/scene_<scene_id>/pred/`. To compute **Collision-Free AP** and
+the collision rate, run the standalone evaluator:
+
+```bash
+cd path_to_openvino_hggd
+conda activate hggd_intel
+
+python evaluate.py \
+  --scene-path /path/to/graspnet \
+  --pred-dir output/scene_100/pred \
+  --scene-l 100 --scene-r 101
+```
+
+The evaluator runs the standard GraspNet grasp pipeline (NMS, object
+assignment, gripper-box collision and empty-grasp test, AP aggregation)
+and counts a grasp as a success when it does not collide with any scene
+point and is not "empty" (at least 10 object points held between the
+fingers). It is fully standalone (numpy + open3d + transforms3d +
+grasp-nms) and only needs `--scene-path` to contain `scenes/` and
+`models/%03d/nontextured.ply` (no textured models required);
+`--scene-path` is the same `/path/to/graspnet` passed to `run.sh`.
+Results are printed to the console and saved to `eval_result_cf.npy`
+next to the prediction directory.
+
+Useful options: `--camera realsense` (default), `--top-k 50`,
+`--proc N` (worker processes), `--result-path /custom/location.npy`.
+
+Note: CF-AP is **not comparable** to the AP / AP@0.8 / AP@0.4 reported
+in the HGGD paper (it has no friction-coefficient semantics) — treat it
+as a separate metric.
+
+To reproduce the AP / AP@0.8 / AP@0.4 as reported in the HGGD paper,
+use the official evaluation toolkit from `graspnetAPI==1.2.11`
+(`pip install --no-deps graspnetAPI==1.2.11`), which evaluates the
+same prediction dumps from this module.
+
+## Output
+
+Results are written under `output/scene_<scene_id>/`.
+
+```
+output/scene_100/
+├── inference.log
+├── eval_result_cf.npy   # written by evaluate.py
+├── logs/
+└── pred/
+```