Baseline
- huggingface
- FP16
- AGX Thor
- Labels
- Image
- Text
- Detection
- Build
- Baseline
- Precision
- FP16
- Runtime
- TensorRT 10.16.2.10
- Input size
- 960×960
- Queries
- Your text, up to 16 tokens
Give OWLv2 a photo and a phrase such as “red mug” or “person on a skateboard”, and it finds the matching objects and returns labelled boxes.
OWLv2 weights and Transformers code are Apache-2.0; the example photos are CC BY 2.0, with sources in examples.json.
Methodology
- Latency: Batch 1, two seconds warmup, ten seconds CUDA-event sampling per engine. Socket median includes decoding, preprocessing and copies, with cached text. Text engine time is reported separately.
- Accuracy: All 5000 COCO val2017 images; all 80 class-name queries with the same a photo of a template. Best class per patch, top 100 patches, no NMS. Same resize and decoder in FP32 and TensorRT; no calibration on val.
- Memory: Engine file size and peak server process RSS are distinct. GPU activation memory is recorded in target.json; RSS is not total GPU allocation.
- Size: the image and text TensorRT engines on disk.
- PyTorch FP32: the upstream checkpoint in PyTorch FP32 with the same resize and decoder, as the accuracy reference.
- Queries: short text queries (16 tokens including “a photo of a”); spatial relations and long referring expressions may be unreliable. COCO class-name evaluation is text-conditioned detection, not proof that every class was unseen in training.