FaceBoxes PyTorch — OpenVINO FP16

This repository distributes the existing OpenVINO IR conversion of FaceBoxes used by Advantech Intel VisionAI face detection. The original PyTorch model is from zisianw/FaceBoxes.PyTorch, as referenced by Open Model Zoo. Advantech did not train the original model.

Files and interface

Download faceboxes-pytorch.xml and faceboxes-pytorch.bin into the same directory, retaining their filenames. This is IR v11; the XML records OpenVINO 2024.6.0 runtime metadata. Artifacts are unchanged from the existing VisionAI payload, without reconversion. FP16 describes compressed stored weights, not the precision of all operations; the input/output tensors are F32.

Tensor	Type	Shape
Input `input.1`	F32	`[1, 3, 1024, 1024]` (NCHW)
Output `boxes`	F32	`[1, 21824, 4]`
Output `scores`	F32	`[1, 21824, 2]`

Expected image channel order is BGR. Use the matching FaceBoxes prior-box decoder and non-maximum suppression. Preprocessing must follow this conversion; the upstream conversion uses BGR means [104, 117, 123]. Result-layer names can include /sink_port_0; the tensor aliases above describe the model's outputs.

Local use

Requires the VisionAI C++ demo, OpenVINO 2026.3 runtime and suitable device driver. From the directory containing the downloaded files:

sha256sum -c SHA256SUMS
export LD_LIBRARY_PATH=/opt/Advantech/EdgeAI/System/Intel/SDK/runtime/openvino_2026_3/lib
/opt/Advantech/EdgeAI/System/Intel/VisionAI/app/bin/object_detection_demo \
  -m ./faceboxes-pytorch.xml -at faceboxes -d CPU -i /path/to/input.mp4 -loop

GPU/NPU compatibility and performance depend on runtime and drivers. No new accuracy evaluation or guaranteed video FPS is claimed. This is a local model artifact repository, not a hosted face-detection service.

Sources and license

The original license is MIT. Its complete license text and copyright notices are retained in LICENSE. SHA256SUMS covers both model files.

Downloads last month: -; Downloads are not tracked for this model. How to track