SCKEMPER/centerface-dynamic

CenterFace — dynamic-input-shape ONNX (RootReal-curated)

A dynamic-input-shape ONNX export of CenterFace (anchor-free face detection + 5-point landmarks), curated by RootReal to fix Forgejo #11856.

Why this exists

The upstream Star-Clouds/CenterFace export (centerface.onnx, MIT) declared a fully static input input.1 = [10, 3, 32, 32] (fixed batch of 10 AND fixed 32×32 spatial dims). ONNX Runtime therefore rejects any real detection tensor (1, 3, H, W):

INVALID_ARGUMENT: Got invalid dimensions for input: input.1
  index: 0 Got: 1 Expected: 10
  index: 2 Got: H Expected: 32
  index: 3 Got: W Expected: 32

The CenterFace graph is purely convolutional (60 Conv / 59 BatchNormalization / 42 Relu / 13 Add / 3 ConvTranspose / 1 Sigmoid — no Reshape/Flatten/Tile that could bake in static shape values), so the static shape is purely an export artifact. This model rewrites input.1 to [batch(dynamic), 3, height(dynamic), width(dynamic)] with zero weight/node change — the canonical equivalent of passing dynamic_axes={'input.1': {0:'batch', 2:'height', 3:'width'}} to torch.onnx.export, without needing PyTorch.

Provenance / reproduction

• Derived from: Star-Clouds/CenterFace models/onnx/centerface.onnx (MIT).
• Upstream sha256: 77e394b51108381b4c4f7b4baf1c64ca9f4aba73e5e803b2636419578913b5fe
• This artifact sha256: 884bf341ce5ba95a30667f86eaa36f24d62041d9d28002eee3bb6926701924bf
• Size: 7,532,788 bytes (362 initializers + 178 nodes, byte-identical to upstream).
• Reproduce: infrastructure/scripts/ml/centerface-rewrite-input-shape-dynamic.py in the RootReal repo regenerates this file byte-for-byte from the upstream MIT source.

Accuracy

WIDER FACE val Easy/Medium/Hard: 0.935 / 0.924 / 0.875 (unchanged from upstream — only the input shape declaration differs; the trained weights are identical).

Migration note

Hosted under a personal namespace (SCKEMPER/centerface-dynamic) because the RootReal Hugging Face org (rootreal/) was not writable at curation time. Migration target: rootreal/centerface (pending org access). The model is a deterministic MIT derivation fully reproducible from the upstream source via the committed script, so re-hosting is trivial.