Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection
Paper • 2303.05499 • Published • 9
Full-precision ONNX export of IDEA-Research/grounding-dino-base for Zuma Innovations InSpec Zuma Ground.
The graph uses the same I/O contract as onnx-community/grounding-dino-tiny-ONNX so InSpec can run Tiny and Base interchangeably.
onnx/model.onnx (~923 MB) — Grounding DINO-Bvocab.txt — BERT-base-uncased vocab (same family as Tiny)| Name | Shape | Type |
|---|---|---|
pixel_values |
[1, 3, 800, 800] |
float32, ImageNet mean/std |
pixel_mask |
[1, 800, 800] |
int64, 1 = content, 0 = pad |
input_ids |
[1, 256] |
int64 |
token_type_ids |
[1, 256] |
int64 |
attention_mask |
[1, 256] |
int64 |
Captions are lowercased phrases joined by ..
| Name | Shape | Notes |
|---|---|---|
logits |
[1, Q, 256] |
pre-sigmoid token similarities |
pred_boxes |
[1, Q, 4] |
cxcywh normalized 0–1 |
https://huggingface.co/ZumaInnovations/zuma-ground-base-onnx/resolve/main/onnx/model.onnx
https://huggingface.co/ZumaInnovations/zuma-ground-base-onnx/resolve/main/vocab.txt
Base model
IDEA-Research/grounding-dino-base