sam2-lite

SAM 2.1 video object segmentation 2.13× faster on CPU: SAM 2.1 Hiera-tiny with its image encoder distilled into a MobileNetV4 backbone and its memory attention distilled to attend to 3 memory frames instead of 7. Click an object in the first frame and it is tracked through the video.

Code, training and evaluation: https://github.com/pablopigue/sam2-lite

Results

DAVIS 2017 val (30 videos, semi-supervised, J&F), per-frame latency of the full video pipeline (median, fp32, 13th Gen Intel(R) Core(TM) i7-13620H).

	SAM 2.1 Hiera-T (teacher)	sam2-lite @1024	sam2-lite-mobile @576
J&F	89.1	84.4	77.1
CPU 6 threads (ms/frame)	1956	920	–
CPU 2 threads (ms/frame)	4949	2270	398
Model size, fp32 (MB)	149	74	74

sam2-lite-mobile is the same weights at a 576×576 input; it was measured in a different session from the other latency numbers (~10 % run-to-run variation). 2 CPU threads = a modest 2-core CPU.

ONNX Runtime encoder

The image encoder is also provided as ONNX (one file per resolution). PyTorch → ONNX Runtime, all measured in one later session (hence e.g. 892 instead of 920 ms for the same PyTorch model: ~5-10 % varies between sessions, so only same-session numbers are compared):

	Encoder (ms)	Full pipeline (ms)
1024, CPU 6 threads	224 → 142 (1.58×)	892 → 830 (1.07×)
1024, CPU 2 threads	434 → 295 (1.47×)	2236 → 2075 (1.08×)
576, CPU 6 threads	44 → 40 (1.11×)	159 → 151 (1.06×)
576, CPU 2 threads	96 → 75 (1.28×)	392 → 373 (1.05×)

The encoder is faster but the pipeline gains little: the memory attention, which still runs in PyTorch, dominates each frame. Quality is unchanged (5 val videos at 576: J&F 82.35 with PyTorch, 82.35 with ONNX Runtime).

Files

• model.safetensors + config.yaml: the whole tracker (student encoder, fine-tuned memory attention, SAM 2.1-tiny memory encoder and mask decoder) and what is needed to rebuild it.
• student_encoder_r1024.onnx, student_encoder_r576.onnx: the image encoder for ONNX Runtime (CPU).
• LICENSE, LICENSE-APACHE-2.0, NOTICE: see License.

Usage

git clone https://github.com/pablopigue/sam2-lite && cd sam2-lite && uv sync

from huggingface_hub import snapshot_download
from sam2lite.export.bundle import load_bundle

path = snapshot_download("Pablopigue/sam2-lite")
predictor = load_bundle(path)  # sam2-lite @1024
mobile = load_bundle(
    path, image_size=576, onnx=f"{path}/student_encoder_r576.onnx"
)  # sam2-lite-mobile, ONNX encoder
# Then use it like SAM 2's video predictor: init_state, add_new_points_or_box, propagate_in_video.

Training

• Image encoder: MobileNetV4-Conv-Medium (timm, ImageNet-1k pretrained) + SAM 2's FPN neck, trained to reproduce the teacher encoder's multi-scale features (MSE per FPN level, FitNets-style), on ~3.8k unlabeled frames from 54 DAVIS 2017 train videos. The rest of SAM 2.1 is reused unchanged.
• Memory attention: a copy of SAM 2.1's, fine-tuned to reproduce the original's output (7 memory frames) while attending to 3 (stride 4), on DAVIS 2017 train clips.
• DAVIS 2017 val was used only for the final evaluation.

Limitations

• Accuracy drops by 4.7 J&F points at 1024; a few small or close-up objects explain most of the drop.
• sam2-lite-mobile misses its own quality target (J&F ≥ 80) while meeting its latency target (≤ 600 ms with 2 threads).
• Distilled on a small dataset (DAVIS 2017 train only); the experiments point at the data as the current bottleneck.
• Latency was measured on one laptop CPU; numbers on other hardware will differ.

License

sam2-lite is an independent project; it is not affiliated with, endorsed by or sponsored by Meta. "SAM 2" refers to the original model by Meta FAIR, on which this work is based.

Released under CC BY-NC 4.0 for non-commercial research use only (LICENSE); NOTICE details every part:

• It contains SAM 2.1 weights by Meta Platforms, Inc., licensed under the Apache License 2.0 (LICENSE-APACHE-2.0). Modified: the memory attention (fine-tuned to attend to 3 memory frames instead of 7) and the memory temporal encodings (3 of the 7 kept). The other SAM 2.1 weights (memory encoder, prompt encoder, mask decoder, object pointers) are unchanged and remain available under the Apache License 2.0.
• The image encoder starts from the timm weights mobilenetv4_conv_medium.e500_r224_in1k (Apache 2.0), modified by distillation. They were pretrained on ImageNet-1k, whose terms allow only non-commercial research and educational use.
• It was distilled on frames of DAVIS 2017, licensed under CC BY-NC 4.0; about half of its sequences come from third-party sources (mostly YouTube) with their own terms. No DAVIS videos, frames or annotations are included.

Citations

@article{ravi2024sam2,
  title   = {SAM 2: Segment Anything in Images and Videos},
  author  = {Ravi, Nikhila and others},
  journal = {arXiv preprint arXiv:2408.00714},
  year    = {2024}
}
@article{pont2017davis,
  title   = {The 2017 DAVIS Challenge on Video Object Segmentation},
  author  = {Pont-Tuset, Jordi and Perazzi, Federico and Caelles, Sergi and Arbel{\'a}ez, Pablo and Sorkine-Hornung, Alexander and Van Gool, Luc},
  journal = {arXiv preprint arXiv:1704.00675},
  year    = {2017}
}
@inproceedings{perazzi2016davis,
  title     = {A Benchmark Dataset and Evaluation Methodology for Video Object Segmentation},
  author    = {Perazzi, Federico and Pont-Tuset, Jordi and McWilliams, Brian and Van Gool, Luc and Gross, Markus and Sorkine-Hornung, Alexander},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2016}
}
@article{qin2024mobilenetv4,
  title   = {MobileNetV4: Universal Models for the Mobile Ecosystem},
  author  = {Qin, Danfeng and others},
  journal = {arXiv preprint arXiv:2404.10518},
  year    = {2024}
}

MLflow runs behind every number: b44efdd1, 62e59f5e, ca473c99, a0b4ebcc, a384f235, b205e7be, af0d9753, 7baca321, 79a6f48c, 005f6d98, 7be93e7d, 4e29d891