Dockerized_GroundingSAM

Dockerized pseudolabeling workflow for prompt-based object detection and segmentation using GroundingDINO + Segment Anything.

This project was built around a warehouse/logistics labeling use case: generate labels from natural-language prompts, filter them geometrically, and export visualization artifacts plus rotated-box annotations.

What it does

• Runs the labeling stack inside Docker
• Uses text prompts to detect target objects in images
• Applies segmentation masks to refine detections
• Filters detections by ROI, area, and IoU
• Writes:
  • labeled visualizations
  • masks
  • synthetic overlays
  • JSON label output

Key files

• pseudolabel.sh — builds and launches the Docker environment
• Dockerfile — container definition
• label_app.py — main labeling application
• gradio_demo/ — demo UI experiments
• utilities/ — filtering, file management, visualization, and bounding-box utilities
• example_images/ — sample input imagery

Quick start

Build the image and launch the container:

./pseudolabel.sh

The launch script currently:

1. builds the Docker image as pseudolabel_app
2. starts a GPU-enabled container with an output volume mounted at /workspace/tool_output

Example invocation inside the container

python label_app.py \
    --image_path '/workspace/Packages2Overlay' \
    --output_path '/workspace/Packages2Overlay_labeled' \
    --confidence_score 0.3 \
    --prompt 'parcel,package,clothing bag,jeans,bag,box,envelope,plastic,white square' \
    --background_path '/workspace/otcempty1.bmp' \
    --maxmin_area 2231850 70000 \
    --max_iou 0.01

Expected environment

This repo assumes a development environment with:

• Docker
• NVIDIA GPU runtime support
• model weights mounted or available inside the container
• GroundingDINO / SAM dependencies installed through the image

Outputs

The pipeline writes artifacts under the configured output directory, including visualized labels, masks, and generated overlay images for inspection.

Related repo

This repo is the more containerized counterpart to:

• aarwitz/GroundingDINO_Autolabeling