run_object_detection.txt

Let's try another vision task that builds on the previous one: Object Detection.

Instead of just classifying the whole image, object detection models identify multiple objects within an image and draw bounding boxes around them, also classifying each detected object.

Prerequisites:

This example uses the same core vision libraries as the previous one. Ensure you have them installed:

Bash

pip install transformers torch Pillow torchvision timm requests
(Note: We added requests to optionally download a sample image).
(Ensure your existing virtual environment is active).

Explicit Model Choice:

We will use facebook/detr-resnet-50. This is a DETR (DEtection TRansformer) model, which applies the transformer architecture to object detection, combined with a standard ResNet-50 backbone for feature extraction. It's trained on the COCO dataset, which includes 91 common object categories (people, cars, dogs, cats, tables, chairs, etc.).

How to Run:

Image Choice:
You can place your own image file (containing common objects like people, cars, furniture, animals) in the same directory as the script and rename it to my_object_detection_image.jpg.
OR, just run the script as is. If it doesn't find my_object_detection_image.jpg, it will attempt to download a sample image (detr_sample_image.jpg) containing cats and a remote control.
Make sure you've run pip install transformers torch Pillow torchvision timm requests in your activated virtual environment.
Open your Ubuntu terminal.
Make sure your virtual environment is activated (source .venv/bin/activate).
Run the script:
Bash

python run_object_detection.py
What to Expect:

First Run: It will download the facebook/detr-resnet-50 model files (DETR models can be several hundred MB) and cache them locally. If you didn't provide your own image, it will also download the sample JPG.
Detection Execution: The model will analyze the image (either yours or the downloaded sample).
Output: It will print a list of detected objects that exceed the confidence threshold (0.9 in this case). For each detected object, it will show:
The predicted label (e.g., "cat", "remote", "couch" for the sample image; or "person", "car", "traffic light" for a street scene).
The confidence score.
The coordinates of the bounding box (xmin, ymin, xmax, ymax) around the detected object in the image (pixel values).


What to Expect:

The script will perform the object detection as before.
New: It will then:
Open the image used for detection.
Draw bright green (lime) rectangles around each detected object (with score > 0.9).
Write the object label and confidence score (e.g., "cat: 0.99") in black text on a green background just above each box.
Save the resulting image with annotations as object_detection_output.jpg in the same directory where you run the script.
You can then open object_detection_output.jpg with any image viewer to see the results visually.