insight-platform / similari Goto Github PK

Following the instructions from https://medium.com/inside-in-sight/high-performance-python-sort-tracker-225c2b507562 resulted in a build failure, with the message maturin failed, the module name must not contain a minus '-' . I edited the pyproject.toml file to change the project name back to similari (no -trackers-rs) and the build worked.

Links on benchmarks page seem to be dead.

No source distributions are available pn PyPi making it impossible to upgrade similari-trackers-rs further than version 0.26.5 using pip install --upgrade and so even with the --no-binary or --no-cache-dir options.

When the boxes are heavily overlapped the features tend to be of a poor quality. This request introduces the requirement to avoid using such features in comparisons or collecting them to the track for future use.

1. Unify IoU and Mahalanobis SORTS to decrease the codebase;
2. Add the capability to pass custom_object_id.

Track multi polygons provided by segmentation models with a simple IoU method.

Track objects described by key points.

Hi,

I would like to implement a custom Tracker using the TrackerAPI, whose signature is

pub trait TrackerAPI<TA, M, OA, E, N>
where
    TA: TrackAttributes<TA, OA>,
    M: ObservationMetric<TA, OA>,
    OA: ObservationAttributes,
    N: ChangeNotifier,
    E: EpochDb,
{ .. }

while the trait is public, the module containing the definition is not trackers.rs:

/// Trait that implements epoch db management
mod epoch_db;

Using the pre-implemented SortAttributesOptions is also not possible as the trait's functions are also private, ie. epoch_db() or next_epoch(), therefor I can't use them in my custom implementation without bringing EpochDb into scope.

Is the EpochDB mod deliberately private or I missing something in the API design? Or could we just:

- mod epoch_db;
+ pub mod epoch_db;

Cheers

Greetings, I would like use the class Point2DKalmanFilter for tracking 2d points on a map, say car/plane movements.
How do I properly handle missing data?

One approach would be to set the covariance matrix to a large number, following 1, 2.

I was unable to find a method to directly modify Point2DKalmanFilterState to accomplish that.

Is this the approach you rely on for missing data? I may open a PR with this functionality added, if needed.

Is there any example code available for using BatchVisualSort in python?

To implement the way to split the objects space based on maximal distance for epochs spent between the last merge and a new observation.

The code crashes:
https://pastebin.com/WSJheVPX

Please update geo="^0.25.1" to avoid recursion limit overflow
rust-lang/rust#110475 (comment)

• Rust;
• Python.

Thanks a lot! Sorry for bothering you! 🙏

Create a filter that can predict the evolution of the vector of independent 2D points.

Would it be possible to publish the latest version on crates.io? Thanks a lot!

Many real-time deep learning inference systems that utilize Gstreamer like Nvidia DeepStream, AMD Vitis, and Intel DLStreamer operate with image batches when doing the inference. Thus the tracker must support the batched processing when the batch elements are sent to the tracker at once and awaiting when it completes. This processing approach is way more efficient than sending every element of the batch to the tracker and awaiting for the tracking prediction to complete.

Following your example from https://medium.com/@kudryavtsev_ia/high-performance-sort-tracker-in-rust-9a1dd18c259c

Under MacOS 12.5 with rust 1.62.1, the example won't build with targets x86_64 and arm64 as there seems to be an issue with Pyo3.

See the attached build log.
build.log

• Add batch sort tracker information
• Add docs for new classes and methods
• Fix doc paths.

I have the same issue as #67 with roughly the same configuration. Because my project is made for multiple hosts configurations, adding a dependency on Python through PyO3 without a very good reason is almost a deal breaker.

Is there any plans to make the Python bindings an optional feature?

With SORT, I think you are supposed to calculate the prediction on all tracked objects before the IOU association is done. The current implementation looks like it only runs the prediction if the IOU matches.

This also means that any idle item doesn't advance correctly with time.

Furthermore, if you are skipping epics, the prediction should be executed for those skipped epics. Otherwise, the bounding box would be stuck in older epics.