nlf 2025 部署笔记

目录

jit部署测试命令

nlf-pipepine

nlf-pipeline 依赖项：

stcnbuf 人体分割，没有sam2好

framepump库报错：

分割算法：stcn.pth

相机姿态估计：

---

jit部署测试命令

python -c "import torch;import torchvision;torch.jit.load('/shared_disk/models/others/nlf/models/nlf_l/nlf_l_multi_0.3.2.torchscript').cuda().eval()"

nlf-pipepine

https://github.com/isarandi/nlf-pipeline

pip install stcnbuf

cl: 命令行 error D8021 :无效的数值参数“/Wno-cpp”

pip install stcnbuf

pip install git+https://github.com/isarandi/nlf-pipeline.git

nlf-pipeline 依赖项：

cameravision-0.3.0.tar.gz

bodycompress-0.2.3.dev0.tar.gz

kornia

stcnbuf 人体分割，没有sam2好

blendipose-0.1.2 bodycompress-0.2.3.dev0 boxlib-0.2.1 bpy-3.6.0 cameravision-0.3.0 ffmpeg-python-0.2.0 framepump-0.1.3 future-1.0.0 nlf-pipeline-0.1.0 pyransac3d-0.6.0 pytorch-minimize-0.0.2 shapely-2.1.1 smplfitter-0.2.10 stcnbuf-0.2.1 yt-dlp-2025.5.22 zstandard-0.23.0

framepump库报错：

python -c "from framepump.videolib import (VideoFrames,get_duration,get_fps,get_reader,get_writer,num_frames,trim_video,video_audio_mux)"

解决方法1 

from framepump.framepump import (
    VideoFrames,
    get_duration,
    get_fps,
    get_reader,
    get_writer,
    num_frames,
    trim_video,
    video_audio_mux,
)

from framepump.video_writing import VideoWriter

解决方法2，不要framepump库。

分割算法：stcn.pth

GitHub - hkchengrex/STCN: [NeurIPS 2021] Rethinking Space-Time Networks with Improved Memory Coverage for Efficient Video Object Segmentation

相机姿态估计：

Now we can estimate the camera motion:

python -m nlf_pipeline.run_slam \
    --droid-model-path="$DATA_ROOT/models/droid.pth" \
    --video-path="$INFERENCE_ROOT/videos_in/${vid}.mp4" \
    --mask-path="$INFERENCE_ROOT/masks_semseg/${vid}_masks.pkl" \
    --output-path="$INFERENCE_ROOT/cameras/${vid}.pkl" \
    --smooth

This saves the camera motion as a pickled list of cameravision Camera objects.

To verify that this step gave reasonable results, you can visualize the camera trajectory:

python -m nlf_pipeline.viz_camtraj --video-id=$vid --camera-view

vertex_subset 

   body_models = ['smpl', 'smplx']
        vertex_subset = {
            k: np.load(f'{DATA_ROOT}/body_models/{k}/vertex_subset_1024.npz')['i_verts'] for k in
            body_models}