This plugin masks objects detected by NVIDIA nvinfer plugin. Fast and smooth since all the masking processes are done with GPU. Originally forked from seieric/gst-dsobjectsmosaic.
Must be a better masking solution than masking with nvdsosd.
Note: This plugin is tested with PeopleSegNet, which is an instance segmentation model. Masking process refers to a object's mask_params of NvDsObjectMeta. Other models (e.g. semantic segmentation models) may not attach mask_params and this wouldn't work as expected.
Note: For Jetson only, not works with dGPU.
- Mask objects with cuda
- Specify class ids for which blur should be applied
- Fast and smooth processing
Property | Meaning | Type and Range |
---|---|---|
min-confidence | Minimum confidence of objects to be masked | Double, 0 to 1 |
class-ids | Class ids of objects for which masking should be applied | Semicolon delimited integer array |
- DeepStream 6.1
- OpenCV4 with CUDA support
If your environment satisfies the requirements, just run following commands.
git clone https://github.com/seieric/gst-dsobjectsmask.git
cd gst-dsobjectsmask
sudo make -j$(nproc) install
This is a brief instruction to test this plugin after installtion.
- Download NVIDIA-AI-IOT/deepstream_tao_apps to
/opt/nvidia/deepstream/deepstream-6.1/sources/apps
. - Download required models for deepstream_tao_apps. Refer to the documentation.
- Build custom parser in
/opt/nvidia/deepstream/deepstream-6.1/sources/apps/deepstream_tao_apps/post_processor
. - Run the sample script
samples/peoplesegnet_test.sh
. - You will get the
peoplesegnet_test.mp4
in your curennt directory.