Take-Away Order Accuracy: Release Notes#

Version history and changelog for Take-Away Order Accuracy.

Version 2026.2.0-rc2 (Latest)#

What’s New#

  • DLStreamer base image updated to intel/dlstreamer:2026.2.0-ubuntu24-rc2.

Published Images#

Image

Tag

intel/order-accuracy-take-away

2026.2.0-rc2

intel/order-accuracy-frame-selector

2026.2.0-rc2

intel/order-accuracy-take-away-ui

2026.2.0-rc2

intel/order-accuracy-take-away-rtsp

2026.2.0-rc2

Version 2026.2.0-rc1#

What’s New#

  • DLStreamer base image updated to intel/dlstreamer:2026.2.0-ubuntu24-rc1.

Published Images#

Image

Tag

intel/order-accuracy-take-away

2026.2.0-rc1

intel/order-accuracy-frame-selector

2026.2.0-rc1

intel/order-accuracy-take-away-ui

2026.2.0-rc1

intel/order-accuracy-take-away-rtsp

2026.2.0-rc1

Version 2026.1.0 (Unreleased)#

What’s New#

  • VLM switched to MiniCPM-V-4.5 (INT4) — Take-Away now uses openbmb/MiniCPM-V-4_5-int4 instead of Qwen/Qwen2.5-VL-7B-Instruct, aligning it with Dine-In so both applications share a single OVMS model instance. VLM_PRECISION defaults to int4.

  • setup_models.sh per-app defaults — --app take-away and --app dine-in both default to MiniCPM-V-4.5 INT4. VLM_PRECISION is now derived from the precision suffix of OVMS_MODEL_NAME, and a mismatch between the two emits a warning.

  • Reasoning disabled by default — MiniCPM-V-4.5 is a hybrid reasoning model; the service now sends chat_template_kwargs.enable_thinking=false (configurable via VLM_ENABLE_THINKING) so the token budget is spent on the answer instead of a <think> block. Any stray <think> output is stripped before parsing.

  • Aspect-preserving image preprocessing — frames are fitted into a 448x448 square canvas (LANCZOS downscale, white padding, light contrast/sharpening) matching the Dine-In pipeline, configurable via VLM_IMAGE_MAX_SIZE. This replaces the previous fixed 512x512 resize, which distorted the aspect ratio.

Fixed#

  • Frames are no longer stretched to 512x512, which distorted product shapes.

  • Detections are no longer truncated by reasoning output, which previously caused missing items and roughly 4x higher VLM latency.

  • setup_models.sh no longer migrates legacy Qwen weights into a directory named after a different model.

Version 2026.0.0 (March 2026)#

General Availability Release

This is the first GA release of Take-Away Order Accuracy, promoted from 2026.0-rc2 with no code changes. All functionality is identical to 2026.0-rc2.

Published Images#

Image

Tag

intel/order-accuracy-take-away

2026.0.0

intel/order-accuracy-frame-selector

2026.0.0

intel/order-accuracy-take-away-ui

2026.0.0

intel/order-accuracy-take-away-rtsp

2026.0.0

Version 2026.0-rc2 (March 2026)#

What’s New#

  • OVMS export scripts updated to OVMS 2026.0 release branch (releases/2026/0); openvino and openvino-tokenizers updated to 2026.0.0rc3

  • YOLO model download added to setup_models.sh — YOLO models are now downloaded automatically during setup

  • Parallel mode VLM scheduler improvements to the VLMScheduler batching logic

  • Frame selector fix — corrected frame selection logic in frame-selector-service

  • Order recall in Gradio UI — added order recall/replay functionality

  • RTSP streaming fix — resolved RTSP stream connection issues in the Gradio UI

  • FastAPI and Starlette version update in the Gradio UI image for security/compatibility

  • Benchmark duration increased (BENCHMARK_DURATION default raised)

  • setup_models.sh simplified — script restructured for clarity

Published Images#

Image

Tag

intel/order-accuracy-take-away

2026.0-rc2

intel/order-accuracy-frame-selector

2026.0-rc2

intel/order-accuracy-take-away-ui

2026.0-rc2

intel/order-accuracy-take-away-rtsp

2026.0-rc1

rtsp-streamer image tag remains 2026.0-rc1 — no changes in this release.

Version 2026.0-rc1 (March 2026)#

Initial Release Candidate

Highlights#

  • AI-Powered Order Validation: Real-time take-away order verification using Qwen2.5-VL-7B Vision Language Model

  • Multi-Station Parallel Processing: Concurrent order validation across multiple stations via RTSP streams

  • Intelligent Frame Selection: YOLO11-based frame selection with OpenVINO™ INT8 inference for optimal VLM input

  • Semantic Matching: Hybrid exact/semantic item matching via dedicated microservice

  • Docker Registry Support: Pre-built images published to intel/ Docker Hub namespace

  • Stream Density Benchmarking: Automated latency-based stream density testing

Published Images#

Image

Tag

Size

intel/order-accuracy-take-away

2026.0-rc1

9.64GB

intel/order-accuracy-frame-selector

2026.0-rc1

1.96GB

intel/order-accuracy-take-away-ui

2026.0-rc1

1.3GB

intel/order-accuracy-take-away-rtsp

2026.0-rc1

227MB

Features#

Core Functionality#

  • Dual Service Mode: Single worker mode for development, parallel worker mode for production

  • VLM Integration: Qwen2.5-VL-7B-Instruct via OpenVINO™ Model Server (OVMS) with GPU acceleration

  • Video Processing: GStreamer-based pipeline with RTSP support and configurable FPS

  • Frame Selection: YOLO11 nano model with OpenVINO™ INT8 inference for hand/object detection and frame filtering

  • Semantic Matching: Hybrid exact/semantic item matching with configurable similarity threshold

  • EasyOCR Integration: Order number detection from video frames

Architecture#

  • Station Workers: Production-ready multi-process workers with per-station isolation

  • VLM Scheduler: Time-window batching for throughput optimization

  • 2PC Pipeline Sync: Two-phase commit synchronization between RTSP streamer and processing pipelines

  • Circuit Breaker: Resilient RTSP connectivity with auto-recovery

  • Exponential Backoff: Configurable retry with jitter for transient failures

User Interface#

  • Gradio UI: Web-based interface for video upload and order validation

  • REST API: FastAPI-based endpoints with OpenAPI documentation

  • MinIO Integration: S3-compatible storage for frames and results

Build & Deployment#

  • Registry Mode: make build pulls pre-built images from Docker Hub

  • Local Build Mode: make build REGISTRY=false builds all images locally from source

  • OVMS Auto-Config: graph.pbtxt auto-generated from config.json graph_options for tester-friendly tuning

  • Model Setup Script: setup_models.sh handles VLM model download, EasyOCR model download, and graph configuration

Benchmarking#

  • Stream Density Test: make benchmark-stream-density — automated latency-based stream scaling

  • Fixed Workers Benchmark: make benchmark-oa — throughput testing with configurable workers

  • Single Video Benchmark: make benchmark — end-to-end latency testing

  • VLM Metrics Logger: Detailed performance metrics collection and consolidation

Components#

Component

Image

Description

Order Accuracy Service

intel/order-accuracy-take-away

Core orchestration, GStreamer pipelines, VLM scheduling

Frame Selector

intel/order-accuracy-frame-selector

YOLO11 OpenVINO™ INT8 frame selection

Gradio UI

intel/order-accuracy-take-away-ui

Web interface for order validation

RTSP Streamer

intel/order-accuracy-take-away-rtsp

Video-to-RTSP stream conversion with 2PC sync

OVMS VLM

openvino/model_server:latest-gpu

Qwen2.5-VL-7B model serving

Semantic Service

intel/semantic-search-agent:1.0.0

Semantic text matching microservice

MinIO

minio/minio

S3-compatible object storage

Configuration Defaults#

Variable

Default

Description

BENCHMARK_TARGET_LATENCY_MS

25000

Target latency threshold (25s)

BENCHMARK_INIT_DURATION

10

Warmup time (seconds)

OCR_WARMUP_FRAMES

2

Frames before OCR ready signal

REGISTRY

true

Pull from registry; set false to build locally

Known Issues#

  1. RTSP Reconnection Delay: Initial RTSP connection may take 5-10 seconds

  2. Large Video Upload: Videos >500MB may timeout on slow connections

  3. Order Accuracy Image Size: 9.64GB due to torch+CUDA dependencies (required by EasyOCR)