^{Back | Next |}^Contents
^{Transfer Learning - Object Detection}

Collecting your own Detection Datasets

The previously used camera-capture tool can also label object detection datasets from live video:

When the Dataset Type drop-down is in Detection mode, the tool creates datasets in Pascal VOC format (which is supported during training).

note: if you want to label a set of images that you already have (as opposed to capturing them from camera), try using a tool like CVAT and export the dataset in Pascal VOC format. Then create a labels.txt in the dataset with the names of each of your object classes.

Creating the Label File

Under jetson-inference/python/training/detection/ssd/data, create an empty directory for storing your dataset and a text file that will define the class labels (usually called labels.txt). The label file contains one class label per line, for example:

Water
Nalgene
Coke
Diet Coke
Ginger ale

If you're using the container, you'll want to store your dataset in a Mounted Directory like above, so it's saved after the container shuts down.

Launching the Tool

The camera-capture tool accepts the same input URI's on the command line that are found on the Camera Streaming and Multimedia page.

Below are some example commands for launching the tool:

$ camera-capture csi://0       # using default MIPI CSI camera
$ camera-capture /dev/video0   # using V4L2 camera /dev/video0

note: for example cameras to use, see these sections of the Jetson Wiki:
             - Nano:  https://eLinux.org/Jetson_Nano#Cameras
             - Xavier: https://eLinux.org/Jetson_AGX_Xavier#Ecosystem_Products_.26_Cameras
             - TX1/TX2: developer kits include an onboard MIPI CSI sensor module (0V5693)

Collecting Data

Below is the Data Capture Control window, after the Dataset Type drop-down has been set to Detection mode (do this first).

Then, open the dataset path and class labels that you created. The Freeze/Edit and Save buttons will then become active.

Position the camera at the object(s) in your scene, and click the Freeze/Edit button (or press the spacebar). The live camera view will then be 'frozen' and you will be able to draw bounding boxes over the objects. You can then select the appropriate object class for each bounding box in the grid table in the control window. When you are done labeling the image, click the depressed Freeze/Edit button again to save the data and unfreeze the camera view for the next image.

Other widgets in the control window include:

Save on Unfreeze - automatically save the data when Freeze/Edit is unfreezed
Clear on Unfreeze - automatically remove the previous bounding boxes on unfreeze
Merge Sets - save the same data across the train, val, and test sets
Current Set - select from train/val/test sets
- for object detection, you need at least train and test sets
- although if you check Merge Sets, the data will be replicated as train, val, and test
JPEG Quality - control the encoding quality and disk size of the saved images

It's important that your data is collected from varying object orientations, camera viewpoints, lighting conditions, and ideally with different backgrounds to create a model that is robust to noise and changes in environment. If you find that you're model isn't performing as well as you'd like, try adding more training data and playing around with the conditions.

Training your Model

When you've collected a bunch of data, then you can try training a model on it using the same train_ssd.py script. The training process is the same as the previous example, with the exception that the --dataset-type=voc and --data=<PATH> arguments should be set:

$ cd jetson-inference/python/training/detection/ssd
$ python3 train_ssd.py --dataset-type=voc --data=data/<YOUR-DATASET> --model-dir=models/<YOUR-MODEL>

note: if you run out of memory or your process is "killed" during training, try Mounting SWAP and Disabling the Desktop GUI.
to save memory, you can also reduce the --batch-size (default 4) and --workers (default 2)

Like before, after training you'll need to convert your PyTorch model to ONNX:

$ python3 onnx_export.py --model-dir=models/<YOUR-MODEL>

The converted model will then be saved under <YOUR-MODEL>/ssd-mobilenet.onnx, which you can then load with the detectnet programs like we did in the previous examples:

NET=models/<YOUR-MODEL>

detectnet --model=$NET/ssd-mobilenet.onnx --labels=$NET/labels.txt \
          --input-blob=input_0 --output-cvg=scores --output-bbox=boxes \
            csi://0

note: it's important to run inference with the labels file that gets generated to your model directory, and not the one that you originally created for your dataset. This is because a BACKGROUND class gets added to the class labels by train_ssd.py and saved to the model directory (which the trained model expects to use).

If you need to, go back and collect more training data and re-train your model again. You can restart again and pick up where you left off using the --resume argument (run python3 train_ssd.py --help for more info). Remember to re-export the model to ONNX after re-training.

Next | WebApp Frameworks
Back | Re-training SSD-Mobilenet

^{© 2016-2020 NVIDIA |}^{Table of Contents}