Skip to content

Release Notes - CVEDIA-RT 2023.1.0

Solutions

Security

  • Trigger and Zone name included in the events
  • Motion guided detection has been improved to merge attention areas
  • Ability to get classified motion regions
  • Zones and Tripwires now return information about triggered objects
  • Improved tripwire crossing results for night-time low-fps footage
  • Fixed an issue where a zone would stop counting after the video loops
  • Tuned the thermal instance where the tripwire was counting the bike but not the person
  • Added option to export events in json without crops on separate files
  • Added a mixed model to the detector variant
  • Disabled 'filter edge detections' when not using multiple inference regions

Playground

We added a powerful tool for testing and comparing different neural networks, whether you are using predefined models from CVEDIA's ModelForge or your own custom networks in ONNX format.

With its easy-to-use interface, you can configure the networks to be tested, visualize the results, and compare the outputs of two networks. You can test and compare object detectors, classifiers, or 3D bounding box regressors.

Click on the (i) icon beside the Playground solution to access the documentation.

Basic Detection

  • Added option to export detecion in json without crops on separate files

Intelligent Transporting Systems

  • Added the ability to export tracks with vehicle classes

Other improvements

  • Validated solutions using Remote UI on Amba
  • Solutions have been updated to use the new gstreamer generic pipeline
  • Improved accuracy for ground use-cases by using the best ground detector model

Models improvements

  • Detection of people from aerial perspective
  • Improved the performance of our Person/Vehicle/Animal (PVA) detector to work better across the board

Engine Updates

  • Added the ability to run external custom AI models
  • Improved performance by running models in parallel
  • A HUD feature has been added to draw overlay interface
  • Reduced size of Windows OpenCV linked library
  • Improved documentation for CVEDIA-RT Inference plugin
  • Ability to set default output preset per instance
  • Option to export JSON data from sink as one-liner for json logs

Remote UI

  • NetImGUI has been improved to reduce required bandwidth and limit texture sizes
  • A crashing issue with DX11 has been fixed in NetImGUI

UI / UX

  • New visualizers for Zones have been added (per-zone occupancy, total zone occupancy, and per-zone state)
  • File menu reorganized and renamed to provide better access
  • Addressed issue with jumping play/pause icons causing misclicks
  • Moved playback bar outside of the Input window
  • Improved interface for Zones and added new visualizers
  • Improved Main Menu bar to simplify access to all windows and features
  • Renamed Input window to Analytics window
  • Improved visual style of RT
  • Created a font cache in memory to speed up rendering
  • Changed event date to playback/frame time in the Events visualizer

Lua

  • Improved tracker to return nil if field doesn't exist
  • Added the ability to render the output presets to a buffer using OpenCV
  • Added inferencemanaged.tryLoadNetwork to handle exceptions
  • Added api.utils.NMS function to run the Non-maximum-suppression algorithm to remove redundant bounding boxes

Bug Fixes

  • Fixed X11 error code when installing on Ubuntu 16 system
  • Fixed CVEDIA-RT freezing upon choosing screencap
  • Fixed stopping one instance while another is on focus
  • Fixed video playlist not going to the next video
  • Fixed zones with total hits and current entries interplaced
  • Fixed zones areas always being zero
  • Fixed saving layout/preset with empty name
  • Fixed instance:getConfigValue returning {} instead of nil when path doesn't exist
  • Fixed MP4 export not creating the output folder
  • Fixed Motion stablization trim ratio issue
  • Fixed issue with selecting a zone in the Input panel
  • Fixed issue with modelforge linking to WebClient
  • Fixed RT closing when creating instance with the same name as an existing instance
  • Fixed input:fileReachedEnd() before the last frame
  • Fixed issue with drawing zone in crowd estimation
  • Fixed tracker track/box association issue when iOU is overwritten
  • Fixed zone entries not resetting in certain cases
  • Fixed colors of new zones being lost in the output window
  • Fixed occupancy zone not triggering when a vehicle is inside it
  • Fixed Analytics Setup windows not being resizable when floating
  • Fixed issue with clicking on playback bar not responding
  • Fixed tripwires created through UI no longer showing up in the output
  • Fixed issue with changing detector variant when sensor modality header is closed
  • Fixed selected detector variant not getting honored on start of an instance
  • Fixed crash when starting with object detection turned off
  • Fixed issue with custom detector regions should not include the implicit full-frame region
  • Fixed ground use-cases that were using the aerial model by default causing bad performance on people
  • Fixed the inability to change the tracking lock time in a solution
  • Fixed an issue where turning on Motion Stabilization prevents events from occurring
  • Fixed an issue where seeking in the playbar would always unpause the video playback
  • Fixed the evaluate tool not working with COCO datasets without a 'supercategory' field
  • Fixed a hang when running an instance
  • Fixed an issue where inferencemanaged.loadModel returns nil instead of a boolean as per documentation
  • Fixed issues with Data Visualizer, Analytics setup, and Output not showing any data until the left side panel is opened
  • Fixed the log being spammed with errors from Hungarian.cpp
  • Fixed an issue with not being able to run parking_lot in parallel with another instance
  • Fixed getMatchedTracks returning an empty object the first time the tracks are created
  • Fixed Crowd estimation defaults being incorrect
  • Fixed double detections when using multiple regions
  • Fixed resetting window layout in Windows
  • Fixed CSV and JSON exports
  • Fixed errors in evaluate.exe on NMS models
  • Fixed classification failing on models with batch size higher than 1
  • Fixed files selection filter to be case-insensitive
  • Fixed basic-detection-demo exporting JPG and JSON + crops
  • Fixed an OpenCV error when running two instances
  • Fixed Inspector when selecting bounding box detections
  • Fixed a regression where detection.id is missing
  • Fixed an issue where runinstance would hang on an instance that runs fine in the UI
  • Fixed an issue when exporting multiple sinks in MQTT from a single instance