Please provide the following information when requesting support.
• Hardware (T4/V100/Xavier/Nano/etc) Nano
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) Detectnet_v2
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here) 5.0.0-tf1.15.5
• Training spec file(If have, please share here) train_aug.txt (3.4 KB)
• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.)
Hello,
We are currently focusing on a project involving real-time object detection. To achieve this, we have fine-tuned the DetectNet_v2 model using TAO (we tried vgg19 and resnet 18 as backbones). After fine-tuning, we exported the model to an ONNX file format and converted it into an engine file which was loaded into DeepStream pipeline. The pipeline is functioning correctly, and it successfully detects objects in the video stream. However, we have noticed that the bounding bbox around the detected objects are slightly shifted towards the upper left corner (same pattern).
This is the current pipeline we are running:
gst-launch-1.0 vmbsrc camera=<camera_id> settingsfile=video_demonstrator_final_square.xml ! video/x-raw,format=RGB ! nvvideoconvert compute-hw=GPU ! videoflip method=clockwise ! nvvideoconvert compute-hw=GPU ! m.sink_0 nvstreammux name=m batch-size=1 live-source=0 width=992 height=992 enable-padding=1 gpu-id=0 batched-push-timeout=33333 ! nvinfer config-file-path="machine_vision_sc/config/models/missing-ic-pcb_arts.cfg" ! nvdsanalytics config-file="machine_vision_sc/config/analytics.txt" ! nvmultistreamtiler rows=1 columns=1 width=992 height=992 ! nvdsosd ! nvegltransform ! nveglglessink sync=0
This is the nvinfer configuration file:
deepstream-config.txt (1.0 KB)
When visualizing the tensor board logs, the training images are well interpreted (precision greater than 90%)!
We made sure that aspect ratio (which is 1:1) is always respected:
- Dataset image size: 320x320 (x16)
- Deepstream input size 992x992 (x16)
Any help is appreciated,
Thank you
3 posts - 2 participants