Baseline TensorRT
- embedl/efficienttam-s-512
- FP16
- AGX Thor
- Labels
- Video
- Segmentation
- Build
- Baseline TensorRT
- Precision
- FP16
- Runtime
- TensorRT 10.16.2.10
- Input size
- 512×512
EfficientTAM-S, the small Efficient Track Anything model, follows the objects you mark on a video's first frame through the whole clip. It also segments a photo from a point or a box.
Methodology
- Latency: median GPU time per frame, one tracked object, batch 1, timed with CUDA events around each engine: the image encoder's plus memory attention, decoder and memory encoder's.
- Accuracy: J&F on DAVIS 2017 val, tracked from each object's first-frame mask: 8 sequences, 544 frames, 16 objects, each scored on its own.
- Memory: peak memory while running the model.
- Size: the five TensorRT engines on disk.
- Baseline TensorRT: this build is the baseline: the original checkpoint in FP16 on stock TensorRT, the image encoder's attention kept in FP32.