#machine-learning #similarity #tracking #sort #deep-sort


Machine learning framework for building object trackers and similarity search engines

50 releases (12 breaking)

0.26.2 May 13, 2023
0.25.1 Mar 10, 2023
0.23.0 Dec 6, 2022
0.22.9 Oct 31, 2022
0.19.1 Jul 30, 2022

#56 in Machine learning

Download history 85/week @ 2023-11-01 1/week @ 2023-11-08 86/week @ 2023-11-15 86/week @ 2023-11-22 339/week @ 2023-11-29 87/week @ 2023-12-06 84/week @ 2023-12-13 86/week @ 2023-12-20 86/week @ 2023-12-27 53/week @ 2024-01-03 34/week @ 2024-01-10 68/week @ 2024-01-17 86/week @ 2024-01-24 90/week @ 2024-01-31 3/week @ 2024-02-07 511/week @ 2024-02-14

753 downloads per month


11K SLoC


Rust Rust Rust Rust

Docker Rust 1.67 Docker Python 3.8 Docker Python 3.9 Docker Python 3.10 Docker Python 3.11

⭐ Star us on GitHub — it motivates us a lot!

Similari is a Rust framework with Python bindings that helps build sophisticated tracking systems. With Similari one can develop highly efficient parallelized SORT, DeepSORT, and other sophisticated single observer (e.g. Cam) or multi-observer tracking engines.


The primary purpose of Similari is to provide means to build sophisticated in-memory multiple object tracking engines.

The framework helps build various kinds of tracking and similarity search engines - the simplest one that holds vector features and allows comparing new vectors against the ones kept in the database. More sophisticated engines operate over tracks - a series of observations for the same feature collected during the lifecycle. Such systems are often used in video processing or other systems where the observer receives fuzzy or changing observation results.

Out-of-The-Box Stuff

Similari is a framework to build custom trackers, however it provides certain algorithms as an end-user functionality:

Bounding Box Kalman filter, that predicts rectangular bounding boxes axis-aligned to scene, supports the oriented (rotated) bounding boxes as well.

2D Point Kalman filter, that predicts 2D point motion.

2D Point Vector Kalman filter, that predicts the vector of independent 2D points motion (used in the Keypoint Tracker).

Bounding box clipping, that allows calculating the area of intersection for axis-aligned and oriented (rotated) bounding boxes.

Non-Maximum Suppression (NMS) - filters rectangular bounding boxes co-axial to scene, and supports the oriented bounding boxes.

SORT tracking algorithm (axis-aligned and oriented boxes are supported) - IoU and Mahalanobis distances are supported.

Batch SORT tracking algorithm (axis-aligned and oriented boxes are supported) - IoU and Mahalanobis distances are supported. Batch tracker allows passing multiple scenes to tracker in a single batch and get them back. If the platform supports batching (like Nvidia DeepStream or Intel DL Streamer) the batch tracker is more beneficial to use.

VisualSORT tracking - a DeepSORT-like algorithm (axis-aligned and oriented boxes are supported) - IoU and Mahalanobis distances are supported for positional tracking, euclidean, cosine distances are used for visual tracking on feature vectors.

Batch VisualSORT - batched VisualSORT flavor;

Applicability Notes

Although Similari allows building various tracking and similarity engines, there are competitive tools that sometimes may fit better. The section will explain where it is applicable and what alternatives exist.

Similari fits best for the tracking tasks where objects are described by multiple observations for a certain feature class, not a single feature vector. Also, their behavior is dynamic - you remove them from the index or modify them as often as add new ones. This is a very important point - it is less efficient than tools that work with growing or static object spaces.

Fit: track the person across the room: person ReID, age/gender, and face features are collected multiple times during the tracking and used to merge tracks or provide aggregated results at the end of the track;

Not fit: plagiarism database, when a single document is described by a number (or just one) constant ReID vectors, documents are added but not removed. The task is to find the top X most similar documents to a checked.

If your task looks like Not fit, can use Similari, but you're probably looking for HNSW or NMS implementations:

Similari objects support following features:

Track lifecycle - the object is represented by its lifecycle (track) - it appears, evolves, and disappears. During its lifetime object evolves according to its behavioral properties (attributes, and feature observations).

Observations - Similari assumes that an object is observed by an observer entity that collects its features (uniform vectors) and custom observation attributes (like GPS or screen box position)multiple times. Those features are presented by vectors of float numbers and observation attributes. When the observation happened, the track is updated with gathered features. Future observations are used to find similar tracks in the index and merge them.

Track Attributes - Arbitrary attributes describe additional track properties aside from feature observations. Track attributes is crucial part when you are comparing objects in the wild, because there may be attributes disposition when objects are incompatible, like animal_type that prohibits you from comparing dogs and cats between each other. Another popular use of attributes is a spatial or temporal characteristic of an object, e.g. objects that are situated at distant locations at the same time cannot be compared. Attributes in Similari are dynamic and evolve upon every feature observation addition and when objects are merged. They are used in both distance calculations and compatibility guessing (which decreases compute space by skipping incompatible objects).

If you plan to use Similari to search in a large index, consider object attributes to split the lookup space. If the attributes of the two tracks are not compatible, their distance calculations are skipped.


The Similari is fast. It is usually faster than trackers built with Python and NumPy.

To run visual feature calculations performant the framework uses ultraviolet - the library for fast SIMD computations.

Parallel computations are implemented with index sharding and parallel computations based on a dedicated thread workers pool.

Vector operations performance depends a lot on the optimization level defined for the build. On low or default optimization levels Rust may not use f32 vectorization, so when running benchmarks take care of proper optimization levels configured.

Rust optimizations

Use RUSTFLAGS="-C target-cpu=native" to enable all cpu features like AVX, AVX2, etc. It is beneficial to ultraviolet.

Alternatively you can add build instructions to .cargo/config:

rustflags = "-C target-cpu=native"

Take a look at benchmarks for numbers.

Performance Benchmarks

Some benchmarks numbers are presented here: Benchmarks

You can run your own benchmarks by:

rustup default nightly
cargo bench

Apple Silicone Build Notes

You may need to add following lines into your ~/.cargo/config to build the code on Apple Silicone:

rustflags = "-C target-cpu=native"

# Apple Silicone fix
rustflags = [
    "-C", "link-arg=-undefined",
    "-C", "link-arg=dynamic_lookup",

Python API

Python interface exposes ready-to-use functions and classes of Similari. As for now, the Python interface provides:

  • the Kalman filter for axis-aligned and oriented (rotated) boxes prediction;
  • the Kalman filter for 2D point motion prediction;
  • the 2D Point Vector Kalman filter, that predicts the vector of independent 2D points motion (used in the Keypoint Tracker);
  • NMS (Non-maximum suppression);
  • the Sutherland-Hodgman clipping, intersection area for oriented (rotated) boxes;
  • SORT with IoU and Mahalanobis metric;
  • BatchSORT with IoU and Mahalanobis metric;
  • VisualSORT - DeepSORT-like tracker with euclidean/cosine metric for visual features and IoU/Mahalanobis metric for positional tracking (VisualSort).
  • BatchVisualSORT - batched VisualSORT flavor;

Python API classes and functions can be explored in the python documentation and tiny examples provided.

There is also MOTChallenge evaluation kit provided which you can use to simply evaluate trackers performance and metrics.

Install Python API from PyPi

Please, keep in mind that the PyPi package is built to conform broad range of platforms, so it may not be as fast as the one you build locally for your platform (see the following sections).


  • Linux: X86_64, ARM64, ARMv7;
  • Windows: X86_64;
  • MacOS: X86_64, ARM64.
pip3 install similari-trackers-rs

Build Python API in Docker

You can build the wheel in the Docker and if you want to install it in the host system, copy the resulting package to the host system as demonstrated by the following examples.

Rust 1.67 Base Image

If you use other rust libraries you may find it beneficial to build with base Rust container (and Python 3.8):

docker build -t similari_py -f docker/rust_1.67/Dockerfile .

# optional: copy and install to host system
docker run --rm -it -v $(pwd)/distfiles:/tmp similari_py cp -R /opt/dist /tmp
pip3 install --force-reinstall distfiles/dist/*.whl

Python 3.8 Base Image

Python 3.8 is still a very frequently used. Here is how to build Similari with it:

docker build -t similari_py -f docker/python_3.8/Dockerfile .

# optional: copy and install to host system
docker run --rm -it -v $(pwd)/distfiles:/tmp similari_py cp -R /opt/dist /tmp
pip3 install --force-reinstall distfiles/dist/*.whl

Python 3.10 Base Image

If you use the most recent Python environment, you can build with base Python container:

docker build -t similari_py -f docker/python_3.10/Dockerfile .

# optional: copy and install to host system
docker run --rm -it -v $(pwd)/distfiles:/tmp similari_py cp -R /opt/dist /tmp
pip3 install --force-reinstall distfiles/dist/*.whl

NOTE: If you are getting the pip3 error like:

ERROR: similari_py-0.22.5-cp38-cp38-manylinux_2_28_x86_64.whl is not a supported wheel on this platform.

It means that the Python version in the host system doesn't match to the one that is in the image used to build the wheel.

Build Python API in Host System

Linux Instruction

  1. Install up-to-date Rust toolkit:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
source $HOME/.cargo/env
rustup update
  1. Install build-essential tools apt install build-essential -y.

  2. Install Python3 (>= 3.8) and the development files (python3-dev).

  3. Install Maturin:

pip3 install --upgrade maturin~=0.13
  1. Not in VENV. Build the python module:
RUSTFLAGS=" -C target-cpu=native -C opt-level=3" maturin build --release --out dist
pip3 install --force-reinstall dist/*.whl
  1. In VENV. Build the python module:
RUSTFLAGS=" -C target-cpu=native -C opt-level=3" maturin develop
  1. Usage examples are located at python.
  2. MOT Challenge Docker image for Similari trackers and conventional trackers is here. You can easily build all-in-one Docker image and try ours trackers.

Manuals and Articles

Collected articles about how the Similari can be used to solve specific problems.


Usage Examples

Take a look at samples in the repo:


~340K SLoC