{"id":82470,"date":"2023-08-10T17:47:07","date_gmt":"2023-08-10T09:47:07","guid":{"rendered":"https:\/\/www.seeedstudio.com\/blog\/?p=82470"},"modified":"2023-09-08T09:57:56","modified_gmt":"2023-09-08T01:57:56","slug":"nvidia-jetson-with-autodistill-deploy-computer-vision-with-efficient-auto-labeling-method","status":"publish","type":"post","link":"https:\/\/www.seeedstudio.com\/blog\/2023\/08\/10\/nvidia-jetson-with-autodistill-deploy-computer-vision-with-efficient-auto-labeling-method\/","title":{"rendered":"NVIDIA Jetson with Autodistill: Deploy Computer Vision with Efficient Auto-labeling Method"},"content":{"rendered":"\n

Before releasing the power of the Jetson Edge device for AI inferencing and processing, there’s always a key issue with time-consuming effort and much resources input on it, which is dataset labeling preparation. Traditionally, annotating image data in a manual way could be accurate for model training but should be within a long time process. Autodistill with Grounded-SAM and YOLOv8 can help you simplify the process by automatically labeling training data for edge-developing scenarios.<\/p>\n\n\n

\n
\"\"<\/figure><\/div>\n\n\n

With Roboflow’s Colab guidance<\/a>, it’s quite easy to get started with the annotation part and build your own PyTorch model to be ready for edge deployment.<\/p>\n\n\n\n

Dataset Preparation<\/h2>\n\n\n\n

After checking the GPU environment and finishing all necessary packages installations, we can create images and videos directories to store data sources and pre-annotated images:<\/p>\n\n\n

\n
\"\"<\/figure><\/div>\n\n\n

In real-life applications, we usually want to analyze some pinpoints from the video streaming rather than a huge volume of images waiting for us there as a dataset. For this annotation part, I prepared 10 videos ready to use. The first thing we need to do is convert these videos into images by saving every 10th frame from each video. Then I set the first three videos should be used in the testing part to evaluate the model, and the other seven were used for the training part.<\/p>\n\n\n\n

import supervision as sv\nfrom tqdm.notebook import tqdm\n\nvideo_paths = sv.list_files_with_extensions(\n    directory=VIDEO_DIR_PATH,\n    extensions=[\"mov\", \"mp4\"])\n\nTEST_VIDEO_PATHS, TRAIN_VIDEO_PATHS = video_paths[:2], video_paths[2:]\n\nfor video_path in tqdm(TRAIN_VIDEO_PATHS):\n    video_name = video_path.stem\n    image_name_pattern = video_name + \"-{:05d}.png\"\n    with sv.ImageSink(target_dir_path=IMAGE_DIR_PATH, image_name_pattern=image_name_pattern) as sink:\n        for image in sv.get_video_frames_generator(source_path=str(video_path), stride=FRAME_STRIDE):\n            sink.save_image(image=image)<\/code><\/pre>\n\n\n\n
Autolabel Image<\/h2>\n\n\n\n
When coming to the defining ontology part for labeling, the thing is to figure out how the Base Model is prompted, how it is reflected on your dataset, and what object your target model will detect. I assume the prompt name for each object in Roboflow’s original data library should be common words, so I used “car” and it indeed worked.<\/p>\n\n\n\n
from autodistill.detection import CaptionOntology\n\nontology=CaptionOntology({\n    \"car\": \"vehicle\"\n})<\/code><\/pre>\n\n\n\n
However, if you are confused about using the proper prompt, such as you might have several options like “trash”, “rubbish”, and “waste”, that makes you get no idea which one is the best prompt for labeling our data, check out the blog How to Evaluate Autodistill Prompts with CVevals<\/a> to give it a try, comparing the performance among similar prompt on the dataset, and pick up the best one to use.<\/p>\n\n\n\n
After running the process, we’ll get the sample output with labels on the sample images there.<\/p>\n\n\n
\n
\"\"<\/figure><\/div>\n\n\n
Benefits and Limitation<\/h2>\n\n\n\n
Autodistill provides significant benefits for your computer vision applications:<\/p>\n\n\n\n