- yunyangx/efficient-track-anything
- FP16
- AGX Thor
- Labels
- Video
- Segmentation
- Build
- Baseline
- Precision
- FP16
- Runtime
- TensorRT 10.16.2.10
- Input size
- 512×512
- Prompts
- Points or a box per object, up to 3 objects
Track anything in a video: click a point or draw a box on the first frame, and EfficientTAM follows the object through the clip. A photo works too, as promptable segmentation.
Upstream checkpoint efficienttam_s_512x512.pt at 9bdd8ab585b19ef95f9c9ed847ac9478301890b4 and code at abcd061ebd3cc6e7527d152d75b890126aaa53f6, Apache-2.0; full licence in LICENSE-EFFICIENTTAM. Modified by Embedl: static five-graph export with a fixed memory bank, TensorRT compilation, a C++ tracking loop and serving. DAVIS 2017 data is used for evaluation only and is not part of the model.
Methodology
- Accuracy: DAVIS 2017 val, 8 fixed sequences (544 frames, 16 objects), semi-supervised from the first frame's mask, every object scored on its own (J, F as davis2017-evaluation; first and last frame excluded). The engines run through libetam exactly as served; PyTorch FP32 is upstream's video predictor on the same frames. A release gate, not a benchmark result.
- Throughput: CUDA events around each engine through libetam, batch one, one object; tracking FPS is 1000 / (encoder + memory attention + decoder + memory encoder) medians over the evaluation frames. The socket figure includes MP4 decoding, mask PNGs and JSON.
- Memory: Peak RSS of a fresh serving process after the self-test; on a Jetson GPU memory is shared with the CPU and not all of it appears in RSS.
- Serving: serve.py imports only the Python standard library and Pillow (checked with NumPy and PyTorch blocked); the tracker, memory bank and video decoding are libetam.so, C++ over TensorRT.
- Size: the serialized TensorRT plans of all engines.
Files
Variants
Every build of EfficientTAM-S (512) on AGX Thor.