IDEA-Research/grounding-dino-tiny
Primitive: /extract · Extract ·
Swin
The Grounding DINO model was proposed in Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection by Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, Lei Zhang.
MultimodalBounding boxes
Overview
Hardware: — drives latency, throughput & cost
| Size | 172M params |
|---|---|
| Tasks | /extract |
| License | apache-2.0 |
| Latency | 331 ms |
| Throughput | 0.9 mpix/s |
| Cost | — /1M tok |
Cost is approximate — computed from list GPU prices; your actual price depends on the provider you deploy SIE with.
Extraction
| Output kinds | Bounding Boxes |
|---|---|
| Inputs | text · image |
| Max sequence length | — |
Benchmarks
COCO
Object detection on COCO natural images
Corpus: 5,000 Queries: 5,000
default_limit-1000
Quality
ap 0.4315
ap50 0.5859
ap75 0.4638
ar 100 0.5291
Performance A10G b1 c4
Detect 0.0 mpix/s
Detect p50 43.7s
Performance L4-SPOT b1 c4
Detect 0.9 mpix/s
Detect p50 734.4ms
Performance L4 b1 c4
Detect 0.9 mpix/s
Detect p50 330.7ms
Performance L4 b1 c16
Detect 1.1 mpix/s
Detect p50 263.8ms
default_limit-100
Quality
ap 0.4860
ap50 0.6553
ap75 0.5001
ar 100 0.5593
Performance RTX-4090 b1 c16
Detect 4.0 mpix/s
Detect p50 602.2ms
default
Quality
ap 0.4028
ap50 0.5627
ap75 0.4361
ar 100 0.5211
Performance RTX-PRO-6000-Blackwell-Server-Edition b1 c16
Detect 2.6 mpix/s
Detect p50 113.9ms
Compare (0)Compare →