Unofficial ONNX conversion of DEIM (Intellindust AI Lab) — rtdetrv2-l
Single-file ONNX exporting the raw model outputs (no NMS, no sigmoid, no top-k, no resizing —
all pre/postprocessing stays outside the graph). Unofficial conversion; not endorsed by the
upstream authors. Upstream-reported COCO AP: 54.3 (not re-measured here).
Input contract
| |
|---|
| name | images |
| dtype / layout | float32, NCHW |
| shape | [1, 3, 640, 640] (static) |
Preprocessing: RGB → resize (stretch) to 640×640 → float32 ÷255 → ÷255, no mean/std normalization → HWC→CHW → batch dim.
Output contract
| output | shape | meaning |
|---|
logits | [1, 300, 80] | raw class logits (apply sigmoid yourself) |
pred_boxes | [1, 300, 4] | normalized cxcywh |
80-class COCO (contiguous 0..79), 300 queries, DETR-style set prediction — no NMS.
Artifact
| |
|---|
model.onnx SHA-256 | fada0e0dfc683f65345fb9b1baba085a76ca95e1523ca65731a1f28d1f09a792 (168,908,447 bytes) |
License
