Package Detection¶

At Plumerai we've built the world’s most accurate people, vehicle, animal, and package detection for the edge. We deliver a complete software solution for camera-based object detection. Our models are tiny and fast and detect objects with a high accuracy in any situation.

You can try our AI live in the browser with your webcam.

Technical documentation¶

We have split the technical documentation for our object detection products into the following sections:

Additionally, we provide example usage of the Plumerai Video Intelligence library in the form of a simple OpenCV webcam-to-display pipeline. Documentation and source code is available in a public GitHub repository.

Object detection features¶

Tracking¶

The library allows for tracking each object by providing a unique ID for each detected bounding box. This enables the application to understand the movement of an object from frame to frame. This information can then be used to detect when an object crosses a line or enters a specific area, or determine the direction it is moving into.

Tracking: each bounding box has a different color, representing the unique IDs.

People variety¶

The library accurately detects people in a wide range of situations:

Pose coverage: walking, standing, sitting, standing, and crouching.
Occlusion: where the person is blocked by an object, for example such as a desk or couch.
Diversity: people with different skin tones, ethnicities, genders, and ages.
Clothing and headdress: wearing a wide range of clothing and head coverings, including hats, masks, or a hijab.

A person crouching — Example detections of people in different poses.

A person sitting — Example detections of people in different poses.

A partially occluded person — Example detections of occluded, partially visible people.

Camera¶

The object detection library supports a wide variety of cameras:

Lens: supports a variety of lenses, including wide angle lenses with a 180 degrees field of view.
Resolution: from low resolution like QVGA (320x240) to beyond 4K (3840x2160).
Color: supports both color (RGB) or black and white media.
Aspect ratio: from landscape to portrait, for example 16:9, 1:1, or 9:16.

Lighting¶

The library correctly detects objects under a wide variety of lighting conditions:

Low light: even with noisy or blurry media the detector is accurate.
Night vision: cameras that use NIR LEDs to light up the scene in the dark are supported.

Night vision image — Detections in varying lighting conditions and with IR LED lighting.

Distance¶

Detects objects up close or at a distance:

Close: detects objects that are up close and only partially in view.
Distant: detects objects at more than 20 meters or 65 feet.

Person far away — Examples of detections at a distance.

Mounting¶

A variety of mounting heights and angles are supported:

Standard: mounted at a height of 1-2 meters or 3-6 feet.
High: mounted at a height of 2-5 meters or 6-15 feet.
Low: mounted lower than 1 meter or 3 feet.
Angles: from level to 45 degrees downward.

High angle image — Cameras mounted in different positions and at different angles.

Locations¶

Whatever the location, the object detection performs well:

Indoor: variety of settings such as homes, offices, stores, hospitals, and malls.
Outdoor: variety of locations such as office and home surroundings such as streets, yards, parking lots, sidewalks.

Fixed or moving camera¶

Both fixed and moving cameras are supported:

Fixed: cameras mounted on the wall or embedded into a device such as a display or laptop.
Moving: cameras mounted on a robot or vehicle, or pan-tilt-zoom cameras.

Dataset¶

The software has been trained with over 32 million images and videos to ensure robust performance in a large variety of situations:

Curated: to ensure that all the important use-cases are included, and in the right balance.
Unbiased: we use proprietary data analysis methods to identify specific issues stemming from sampling biases in public datasets.
Varied: our dataset consists of a very large variety of images from all kinds of objects, settings and conditions.
Labeled: we have an extensive methodology to label our datasets to ensure labels are accurate and precise.

Validation¶

Validation: The object detection models have been extensively validated on a diverse range of people, poses, settings, and situations.
Unit tests: Data unit tests guarantee that the model works reliably for specific subsets of our dataset. For example, for people standing far away or for people who are only showing their back.

Privacy¶

Our object detection software ensures that personal data is protected and that privacy is maintained:

Edge compute: our object detection model is small and lightweight, enabling it to run on the device rather than in the cloud where the image data can more easily be accessed or misused.
No image storage: the original image data is not stored on the device, eliminating the possibility of misuse.
No image transmission: since all computations are done at the edge, the images do not need to be transmitted over the internet, eliminating the possibility of interception during transmission.
Distributed: Since the image processing happens on the devices at the edge, instead of in a single location on a central server in the cloud, there’s protection against potential misuse or unauthorized access.