# 套用語義分割至視訊畫面

在執行模型的管線命令前，請完成必要的 [前置條件](https://docs.qualcomm.com/doc/80-70020-15BT/topic/ai-hub-qualcomm-im-sdk.html#prerequisites)。

執行以下命令以確保結果顯示在已連接的顯示器上：

export XDG_RUNTIME_DIR=/dev/socket/weston && export WAYLAND_DISPLAY=wayland-1
    Copy to clipboard

## DeepLabV3-Plus\_MobileNet

DeepLabV3 量化專用於多尺度語義分割，並在各種資料集上進行訓練。

AI Hub 模型以 [此 DeepLabV3-Plus_MobileNet 實作](https://github.com/jfzhang95/pytorch-deeplab-xception) 為基礎。

- 模型： [DeepLabV3-Plus_MobileNet](https://aihub.qualcomm.com/iot/models/deeplabv3_plus_mobilenet)
- 標籤： [voc_labels.txt](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/labels/voc_labels.txt)

gst-launch-1.0 -e --gst-debug=2 \
    filesrc location=/etc/media/video.mp4 ! qtdemux ! queue ! h264parse ! v4l2h264dec capture-io-mode=4 output-io-mode=4 ! video/x-raw,format=NV12 ! queue ! tee name=split \
    split. ! queue ! qtivcomposer name=mixer sink_1::alpha=0.5 ! queue ! waylandsink sync=true fullscreen=true \
    split. ! queue ! qtimlvconverter ! queue ! qtimltflite delegate=external external-delegate-path=libQnnTFLiteDelegate.so \
    external-delegate-options="QNNExternalDelegate,backend_type=htp;" model=/etc/models/deeplabv3_plus_mobilenet_quantized.tflite ! queue ! \
    qtimlvsegmentation module=deeplab-argmax labels=/etc/labels/voc_labels.txt \
    constants="deeplab,q-offsets=<92.0>,q-scales=<0.04518842324614525>;" ! video/x-raw,format=BGRA,width=256,height=144 ! queue ! mixer.
    Copy to clipboard

## FCN-Resnet50

FCN\_ResNet50 是一個量化機器學習模型，可分割 COCO 資料集中的影像。

AI Hub 模型以 [此 FCN-Resnet50 實作](https://github.com/pytorch/vision/blob/main/torchvision/models/segmentation/fcn.py) 為基礎。

- 模型： [FCN-Resnet50](https://aihub.qualcomm.com/iot/models/fcn_resnet50)
- 標籤： [voc_labels.txt](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/labels/voc_labels.txt)

備註

Qualcomm Dragonwing™ RB3 Gen 2 目前不支援此管道。

gst-launch-1.0 -e --gst-debug=2 \
    filesrc location=/etc/media/video.mp4 ! qtdemux ! queue ! h264parse ! v4l2h264dec capture-io-mode=4 output-io-mode=4 ! video/x-raw,format=NV12 ! queue ! tee name=split \
    split. ! queue ! qtivcomposer name=mixer sink_1::alpha=0.5 ! queue ! waylandsink sync=true fullscreen=true \
    split. ! queue ! qtimlvconverter ! queue ! qtimltflite delegate=external external-delegate-path=libQnnTFLiteDelegate.so \
    external-delegate-options="QNNExternalDelegate,backend_type=htp;" model=/etc/models/fcn_resnet50_quantized.tflite ! queue ! \
    qtimlvsegmentation module=deeplab-argmax labels=/etc/labels/voc_labels.txt \
    constants="deeplab,q-offsets=<0.0>,q-scales=<1.0>;" ! video/x-raw,format=BGRA,width=256,height=144 ! queue ! mixer.
    Copy to clipboard

## FFNet-40S

FFNet-40S 是一種「無繁瑣網路」，透過道路、人行道、行人等每像素類別來分割街道場景影像。

此網路透過城市景觀資料集訓練。

AI Hub 模型以 [此 FFNet-40S 實作](https://github.com/Qualcomm-AI-research/FFNet) 為基礎

- 模型： [FFNet-40S](https://aihub.qualcomm.com/iot/models/ffnet_40s)
- 標籤： [voc_labels.txt](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/labels/voc_labels.txt)

gst-launch-1.0 -e --gst-debug=2 \
    filesrc location=/etc/media/video.mp4 ! qtdemux ! queue ! h264parse ! v4l2h264dec capture-io-mode=4 output-io-mode=4 ! video/x-raw,format=NV12 ! queue ! tee name=split \
    split. ! queue ! qtivcomposer name=mixer sink_1::alpha=0.5 ! queue ! waylandsink sync=false fullscreen=true \
    split. ! queue ! qtimlvconverter ! queue ! qtimltflite delegate=external external-delegate-path=libQnnTFLiteDelegate.so \
    external-delegate-options="QNNExternalDelegate,backend_type=htp;" model=/etc/models/ffnet_40s_quantized.tflite ! queue ! \
    qtimlvsegmentation module=deeplab-argmax labels=/etc/labels/voc_labels.txt constants="ffnet,q-offsets=<178.0>,q-scales=<0.31378185749053955>;" ! \
    video/x-raw,format=BGRA,width=256,height=144 ! queue ! mixer.
    Copy to clipboard

## FFNet-54S

FFNet-54S 是一種「無繁瑣網路」，透過道路、人行道、行人等每像素類別來分割街道場景影像。

此網路透過城市景觀資料集訓練。

AI Hub 模型以 [此 FFNet-54S 實作](https://github.com/Qualcomm-AI-research/FFNet) 為基礎。

- 模型： [FFNet-54S](https://aihub.qualcomm.com/iot/models/ffnet_54s)
- 標籤： [voc_labels.txt](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/labels/voc_labels.txt)

gst-launch-1.0 -e --gst-debug=2 \
    filesrc location=/etc/media/video.mp4 ! qtdemux ! queue ! h264parse ! v4l2h264dec capture-io-mode=4 output-io-mode=4 ! video/x-raw,format=NV12 ! queue ! tee name=split \
    split. ! queue ! qtivcomposer name=mixer sink_1::alpha=0.5 ! queue ! waylandsink sync=false fullscreen=true \
    split. ! queue ! qtimlvconverter ! queue ! qtimltflite delegate=external external-delegate-path=libQnnTFLiteDelegate.so \
    external-delegate-options="QNNExternalDelegate,backend_type=htp;" model=/etc/models/ffnet_54s_quantized.tflite ! queue ! \
    qtimlvsegmentation module=deeplab-argmax labels=/etc/labels/voc_labels.txt constants="ffnet,q-offsets=<178.0>,q-scales=<0.2929433584213257>;" ! \
    video/x-raw,format=BGRA,width=256,height=144 ! queue ! mixer.
    Copy to clipboard

## FFNet-78S

FFNet-78S 是一種「無繁瑣網路」，透過道路、人行道、行人等每像素類別來分割街道場景影像。

此網路透過城市景觀資料集訓練。

AI Hub 模型以 [此 FFNet-78S 實作](https://github.com/Qualcomm-AI-research/FFNet) 為基礎。

- 模型： [FFNet-78S](https://aihub.qualcomm.com/iot/models/ffnet_78s)
- 標籤： [voc_labels.txt](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/labels/voc_labels.txt)

gst-launch-1.0 -e --gst-debug=2 \
    filesrc location=/etc/media/video.mp4 ! qtdemux ! queue ! h264parse ! v4l2h264dec capture-io-mode=4 output-io-mode=4 ! video/x-raw,format=NV12 ! queue ! tee name=split \
    split. ! queue ! qtivcomposer name=mixer sink_1::alpha=0.5 ! queue ! waylandsink sync=false fullscreen=true \
    split. ! queue ! qtimlvconverter ! queue ! qtimltflite delegate=external external-delegate-path=libQnnTFLiteDelegate.so \
    external-delegate-options="QNNExternalDelegate,backend_type=htp;" model=/etc/models/ffnet_78s_quantized.tflite ! queue ! \
    qtimlvsegmentation module=deeplab-argmax labels=/etc/labels/voc_labels.txt constants="ffnet,q-offsets=<171.0>,q-scales=<0.3849360942840576>;" ! \
    video/x-raw,format=BGRA,width=256,height=144 ! queue ! mixer.
    Copy to clipboard

Last Published: Dec 23, 2025

[Previous Topic
偵測物件](https://docs.qualcomm.com/bundle/publicresource/80-70020-15BT/topics/object-detection.md) [Next Topic
透過超解析度提升影像畫質](https://docs.qualcomm.com/bundle/publicresource/80-70020-15BT/topics/super-resolution.md)