Model Card for Grounding DINO SwinT OGC (IB-Robot)

Grounding DINO with Swin-Tiny backbone and the OGC checkpoint (groundingdino_swint_ogc.pth, sha256 3b3ca2563c77c69f651d7bd133e97139c186df06231157a64c507099c52bc799) for text-conditioned zero-shot object detection, packaged for the IB-Robot framework with three deployments sharing one raw contract.

Deployments

deployment	backend	artifacts	notes
`ascend_310p`	Ascend ACL (Ascend310P1)	12 OM modules	fixed 720x1280 / seq8 split graph (text/vision/encoder/proposal/decoder/head)
`torch_cuda`	PyTorch CUDA	— (weights in `assets/`)	host-side, ~0.16 s/inference on RTX 3090
`torch_cpu`	PyTorch CPU	— (weights in `assets/`)	host-side fallback (~3.7 s/inference)

Shared raw contract: tensor_model/grounding_dino/detect with fixed-shape inputs (image [1,3,720,1280], BERT seq8 token tensors, encoder_tgt constant for the compiled graph) and outputs (pred_logits [1,900,256], pred_boxes [1,900,4]). The Torch deployments ignore encoder_tgt (the model owns its query embeddings) and saturate the ContrastiveEmbed -inf padding columns to -50 to match the compiled artifacts' finite-value contract.

Weights provenance

All three deployments use the byte-identical IDEA-Research official groundingdino_swint_ogc.pth (same sha256 as the local grounded_sam2_swint_ogc Torch bundle); the 310P OM modules were converted from it.

Repository Structure

inference_manifest.json — deployment routing (schema v3)
inference_integrity.json — per-artifact digests
assets/adapter.json — identity + torch_module_loader
assets/groundingdino_swint_ogc.pth — Torch weights (693 MB)
assets/bert-base-uncased/vocab.txt, assets/encoder_tgt.npy — shared contract assets
artifacts/ascend_310p/*.om — 12 compiled modules

Usage

Host (Torch, via the unified inference runtime):

ros2 run inference_service model_service_node --ros-args \
  -p bundle_path:=models/grounding_dino_swint_seq8_1280x720 \
  -p deployment:=torch_cuda \
  -p adapter_class:=perception_service.model_service_plugins:GroundingDINORawDetectPlugin \
  -p service_type:=ibrobot_msgs/srv/GroundingDetect \
  -p service_endpoint:=/perception/grounding_detect

Board (Ascend 310P): select the ascend_310p deployment (8-role device-linked execution through the unified inference runtime).

License

Code and packaging: Apache-2.0. GroundingDINO weights follow the IDEA-Research upstream license (Apache-2.0).

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