본문으로 건너뛰기
SDK Version: 2.3.3

RELEASE_NOTES

# RELEASE_NOTES

v3.1.1 / 2026-04-21​

Fixed​

  • Fix typo error in document(DX-APP User Manual)

Added​

  • Add License information for third-party models & datasets

v3.1.0 / 2026-04-06​

Changed​

  • Unified 5-layer architecture and design patterns across Python and C++ implementations
    • App Entry point -> yolov5s_sync.cpp : yolov5s_sync.py
    • Runner Pipeline orchestration (Sync/Async) -> sync_detection_runner.hpp : sync_runner.py
    • Factory Per-model component assembly -> yolov5s_factory.hpp : yolov5s_factory.py
    • Component Preprocessor / Postprocessor / Visualizer -> processors/*.hpp : processors/*.py
    • Interface Abstract contracts -> i_factory.hpp, i_processor.hpp : i_factory.py, i_processor.py
  • Consolidated cross-language(Python : c++) common modules and 1:1 mapping structure
  • Modernized run_demo.sh with a 3-stage interactive menu supporting variable AI tasks
  • --model, --image, --video arguments are now optional — when omitted, task-appropriate default sample image/video is automatically selected
  • setup_sample_models.sh migrated to Python-based downloader — supports --list, --dry-run, --category, --models and other granular download options
  • setup.sh now integrates model download options — e.g. setup.sh --models YoloV7 YoloV8S to download specific models
  • run_demo.sh fully redesigned — 18 demo models, 3-stage interactive menu (Task→Mode→Input), unified C++/Python support, --task/--mode/--input CLI arguments for non-interactive usage
  • Display windows now show original resolution frames with auto-sized window (~1/4 screen area)

Added​

  • Supported new Depth Estimation task featuring FastDepth for monocular depth estimation
  • Supported new Image Restoration task featuring DnCNN, Zero-DCE, and ESPCN models
  • Migrated Full DX-Model Zoo encompassing 280 models across 17 taskcategories with 560 C++/Python examples (sync+async)
  • Added yolov8, v9, v10, v11, v12 PPU models and C++/Python examples
  • Implemented https://sdk.deepx.ai manifest-based DX-ModelZoo auto-download system (scripts/download_models.py)
  • Auto-download for models and videos — automatically invokes setup_sample_models.sh when model file is missing, videos via setup_sample_videos.sh
  • Interactive mode for scripts/run_examples.sh — 6-stage menu (Language→Category→Filter→ExecMode→InputType→Options) when run without arguments, with case-insensitive keyword filtering
  • dx_tool.sh run unified with run_examples.sh interactive mode
  • Real-time performance table output during example execution
  • --show-log option for Python examples — controls per-frame detailed log output

v3.0.4 / 2026-03-27​

Changed​

Test Infrastructure Restructuring​

  • tests/common/: extracted shared test constants (constants.py) and utilities (utils.py) from individual test files into a reusable module
  • Unified Visualization Tests: consolidated per-task visualization test scripts into tests/cpp_example/test_visualization.py and tests/python_example/test_visualization.py — auto-discover all sync/async executables and scripts
  • Feature Test Suite: added dedicated test modules for --save / --save-dir (test_save_mode.py), --dump-tensors (test_dump_tensors.py), DXAPP_VERIFY (test_verify.py), --loop (test_multi_loop.py), SIGINT/SIGTERM (test_signal_handling.py)
  • Deleted 4 redundant root-level test scripts (run_e2e_test.sh, run_visualization_test.sh, run_inference_test.sh, run_unit_test.sh) — all functionality covered by run_tc.sh

Documentation Comprehensive Update​

  • README.md: added CLI Reference (full argument tables for C++/Python), Advanced Features section (signal handling, run directory, DXAPP_VERIFY, tensor dump, model config, version compatibility, headless mode), environment variables table, updated module counts and test tree
  • C++ Usage Guide (03): added CLI Arguments table (12 flags), Advanced Features section, updated utility count (6→8)
  • Python Usage Guide (05): added CLI Arguments table (13 flags), Advanced Features section, updated module counts (processors 34→35 description, visualizers 9→10 description)
  • Example Source Structure (11): updated C++ utility count (6→8), Python utility count (6→7), added complete test infrastructure section with 9 test categories, tests/common/ documentation
  • C++ Test README: added visualization tests, feature tests (5 types), test infrastructure section, updated coverage summary
  • Python Test README: expanded to all 14 task directories, added test_visualization.py, tests/common/ reference

Added​

  • tests/common/constants.py — shared test constants (paths, timeouts, patterns)
  • tests/common/utils.py — shared test utilities (discovery, execution, validation)
  • tests/cpp_example/test_visualization.py — unified C++ visualization result tests
  • tests/cpp_example/test_save_mode.py — --save / --save-dir flag tests
  • tests/cpp_example/test_dump_tensors.py — --dump-tensors output tests
  • tests/cpp_example/test_verify.py — DXAPP_VERIFY environment variable tests
  • tests/cpp_example/test_multi_loop.py — --loop repeated execution tests
  • tests/cpp_example/test_signal_handling.py — SIGINT/SIGTERM graceful shutdown tests
  • tests/python_example/test_visualization.py — unified Python visualization result tests

Removed​

  • run_e2e_test.sh — replaced by run_tc.sh --cpp --e2e
  • run_visualization_test.sh — replaced by run_tc.sh --cpp --viz
  • run_inference_test.sh — replaced by run_tc.sh --python
  • run_unit_test.sh — replaced by run_tc.sh --unit

v3.0.3 / 2026-03-13​

Changed​

  • Expanded model support across multiple AI task categories

Fixed​

Post-Processing Bug Fixes (16 models)​

  • Segmentation (bisenetv1, bisenetv2, deeplabv3plusmobilenet): Fixed pre-argmaxed NPU output (uint16 [1,H,W]) being misinterpreted as logits — added heuristic detection for integer dtype / shape[0]==1
  • NanoDet (nanodet_repvgg, nanodet_repvgga1): Fixed degenerate bounding boxes (y2==y1) caused by clipping to image boundary — added zero-area box filter
  • FastDepth (fastdepth_1): Fixed DepthResult not handled by verify_serialize.py — added DepthResult serialization handler
  • YOLOv5Pose PPU (yolov5pose_ppu): Fixed raw logit keypoint confidence (negative values) — applied sigmoid activation
  • YOLOX (yoloxs, yoloxtiny, yolox_l_leaky, yolox_s_leaky, yolox_s_wide_leaky): Fixed zero bounding boxes from raw logit coordinates — implemented standalone YOLOXPostprocessor with grid decode (cx=(cx_raw+grid_x)*stride)
  • YOLOv7 Face (yolov7_face, yolov7s_face, yolov7_w6_face, yolov7_w6_tta_face): Fixed confidence > 1.0 from misread column layout — added auto-dispatch by output column count (16-col vs 21-col) with sigmoid on raw class logit

Added​

Shared Runtime Layer (common/)​

  • C++ (src/cpp_example/common/): Base interfaces (IFactory, IProcessor, IVisualizer, IInputSource), 45 processors, 24 task-specific sync/async runner pairs, 12 visualizers, input source abstraction, config loader, utility
  • Python (src/python_example/common/): Base interfaces (IFactory, IProcessor, IVisualizer, IInputSource), 35 processors, generic SyncRunner/AsyncRunner, 10 visualizers, input source abstraction, ModelConfig loader, utility
  • Both languages share the same 7-module architecture (base/, config/, processors/, runner/, inputs/, visualizers/, utility/) and factory-based delegation pattern

Model Registry System​

  • config/model_registry.json: centralized registry of 280 models with per-model metadata (task, postprocessor, input dimensions, thresholds)
  • scripts/add_model.sh: registry-driven auto-generation of factory files, config.json, and entry-point scripts (4 variants per model)

Numerical Verification Framework​

  • scripts/verify_inference_output.py: 14 task-specific validators for bounding boxes, confidence ranges, class IDs, keypoints, segmentation masks, depth maps, embeddings
  • scripts/inference_verify_rules.json: configurable thresholds per task type
  • common/runner/verify_serialize.py: result-to-JSON serialization for automated comparison
  • scripts/validate_models.sh --numerical: full-pipeline NPU verification for all supported models

New Model Families​

  • DAMOYOLO (5 variants), NanoDet (2), CenterNet, SSD MobileNet V1/V2, YOLOv3, YOLOv6
  • FastDepth, MiDaS (depth estimation)
  • CLIP, ArcFace (embedding)
  • DnCNN, Zero-DCE (image denoising/enhancement)
  • ESPCN (super resolution)
  • Hand Landmark (2 variants)
  • BiSeNet V1/V2, SegFormer (semantic segmentation)
  • YOLOv7Face (4 variants), RetinaFace (face detection)
  • YOLOv26 OBB, Seg, Pose, Cls variants

CI/CD Integration​

  • python-test.yml: automated unit/CLI/integration tests on PR (Python, no NPU required)
  • npu-model-verify.yml: NPU numerical verification pipeline (self-hosted runner with NPU hardware)

v3.0.2 / 2026-02-10​

Changed​

  • Copy of dxrt and vkpkg DLLs into the dx-app/bin directory when building with MSVC.

Fixed​

  • Remove experimental filesystem includes and update float literals in example cpp files for build error on windows
  • Refactor apply_argmax to reduce nesting and fix gcovr warnings

Added​

  • Added vcpkg installation script for windows build.

v3.0.1 / 2026-02-05​

Fixed​

  • Hardcoded attribute size in YOLO post-processing to dynamically adjust based on model output shape

Added​

  • Add yolov26 cls, yolo26 pose, yolo26 seg, yolo26 obb examples

v3.0.0 / 2026-01-02​

Changed​

Major Project Structure Refactoring​

  • Complete overhaul from existing demo applications to example system: To improve user understanding, separated the previously integrated example code by Task (classification, object detection, segmentation, face recognition, pose estimation) / Model (EfficientNet, YOLO, YOLO_PPU, SCRFD, ...) / Inference method (sync, async) / Post-processing (pure python, pybind)
    : Complete removal of legacy C++ demo code in demos/ directory and provision of run_demo.sh and run_demo.bat based on separated examples
    : Transition to new src/cpp_example/ and src/python_example/ structure

Build System Improvements​

  • Improved CMake configuration and enhanced shared library support
  • Updated C++17 and Visual Studio 2022(v143) configuration for Windows build
  • Adjusted DXRT include and link directories for cross-compilation

Complete Reconstruction of C++ / Python Example System​

  • Support for synchronous and asynchronous execution modes
  • Support for various input sources: image, video, camera, RTSP stream
  • Real-time processing mode: Performance measurement without GUI using --no-display option
  • Enhanced performance profiling:
    : Latency measurement for each stage: preprocessing, inference, post-processing
    : E2E(End-to-End) FPS calculation and performance report generation
    : Automatic generation of timestamp-based performance report files

Model Support Expansion​

  • YOLOv10, YOLOv11, YOLOv12 examples added
  • YOLOv8 Segmentation (YOLOv8-seg) support
  • DeepLabv3 segmentation model support
  • PPU (Post-Processing Unit) module integration:
    : YOLOv5, YOLOv7 PPU version support
    : SCRFD PPU version support
    : Both Python and C++ examples provided

Documentation Improvements​

  • Newly written example guides and installation guides
  • Added detailed usage examples and parameter descriptions for each model

Fixed​

Code Quality Improvements​

  • Added try-catch error handling to all projects
  • Improved std::exception handling and throw std::invalid_argument when layer requirements are not met
  • Removed using namespace std usage and improved code clarity with explicit std:: usage
  • Improved parameter handling and frame processing logic
  • Enhanced argument validation and error messages

Input Processing Improvements​

  • Set cv2.CAP_PROP_BUFFERSIZE (buffer size 1) for camera and RTSP speed improvement
  • Fixed input_tensor passing to maintain memory reference until asynchronous inference completes

Added​

Post-processing Library (dx_postprocess)​

  • Pybind11-based Python binding:
    : Provides Python binding for C++ post-processing functions
    : Automatically installs to current Python execution environment

Multi-channel Processing Support​

  • C++ YOLOv5s multi-channel processing: Multi-channel support using frame provider
  • Added multi-input source examples
  • Enhanced multi-model image inference using preprocessing threads

Test Infrastructure Construction​

  • Pytest-based integrated test system:
    : Automated testing for all Python examples
    : Achieved code coverage of 93.65% or higher
    : E2E(End-to-End) test framework
    : Includes all model tests for classification, object detection, segmentation, and pose estimation
  • Added .coveragerc file for code coverage configuration
  • Support for display mode and E2E mode testing

New Examples and Features​

  • Classification Models:
    : EfficientNet example integration
    : ImageNet classification examples (synchronous/asynchronous)

  • Object Detection Models:
    : YOLOv5, YOLOv7, YOLOv8, YOLOv9, YOLOv10(python only), YOLOv11(python only), YOLOv12(python only)
    : YOLOX
    : SCRFD (face detection)
    : YOLOv5-Face
    : YOLOv5-Pose (pose estimation)

  • Segmentation Models:
    : YOLOv8-seg (instance segmentation)
    : DeepLabv3 (semantic segmentation)

  • Pose Estimation:
    : YOLOv5-Pose examples
    : Added skeleton drawing functionality

Removed or Replaced​

Legacy Demo Removal​

  • Complete removal of demos/classification/
  • Complete removal of demos/object_detection/
  • Complete removal of demos/segmentation/
  • Complete removal of demos/pose_estimation/
  • Complete removal of demos/face_recognition/
  • Removal of demos/denoiser/
  • Removal of demos/dncnn_yolo/
  • Removal of demos/object_det_and_seg/
  • Removal of demos/noiseVideoMaker/

Legacy Configuration File Removal​

  • Complete removal of JSON configuration files in example/ directory
  • Removal of example/dx_postprocess/ JSON files
  • Removal of Debian package related files (debian/)
  • Removal of Docker build files (docker/Dockerfile.app.build)

Legacy Code Cleanup​

  • Removal of demo_utils/ directory
  • Removal of duplicate or unused code
  • Removal of old YOLOv5 post-processing files
  • Removal of RISCV64 architecture support

Migration Guide​

Notice for Existing Users​

v3.0.0 is a major update that includes Breaking Changes compared to v2.x.

  • Demo Code: The existing demos/ directory has been completely removed. Please refer to the new examples in src/cpp_example/ and src/python_example/.
  • JSON Configuration Files: JSON files in the existing example/ directory have been removed. Python examples are configured directly through command-line arguments.
  • YOLO Post-processing Type Names: Some have been changed, but aliases are provided for backward compatibility.
  • (1) Refer to Python example documentation
  • (2) Check example code in src/python_example/ or src/cpp_example/ directory
  • (3) Install Python dependencies through requirements.txt
  • (4) Use build scripts (build.sh or build.bat)

Known Issues​

  • When using the PPU model for face detection & pose estimation, dx-compiler v2.1.0 and v2.2.0 does not currently support converting face and pose models to PPU format. This feature will be added in a future release. The PPU models used in the demo were converted using dx-compiler v1.0.0(dx_com v1.60.1).

v2.1.0 / 2025-11-28​

Changed​

  • Enhance build script documentation and usage instructions
  • Update cmake configuration in build.bat to use C++17 and v143 for enhance documentation windows build script(visual studio 2022)
  • Model package updated from version 2.0.0 to 2.1.0 to support PPU models
  • Improved demo script with additional PPU-Demo (1, 4, 6, 8, 11)
  • Added CPU-specific PyTorch wheel source (https://download.pytorch.org/whl/cpu) in templates/python/requirements.txt.

Fixed​

  • Fix Windows MSBuild compilation warnings by replacing implicit type casts with explicit static_cast
  • Improve tensor allocation in imagenet classification example
  • Update numBoxes calculation based on post-processing type in LayerReorder
  • Rename YOLO post-processing types and add aliasing for backward compatibility
  • Add VSCode configuration files for usability
  • Fixed errors that occurred when using VAAPI with camera input
  • Enhanced yolo application to display final FPS even when forcefully terminated during camera input usage
  • Enhance user input handling for run_demo selection with re-prompt loops (invalid input re-asks instead of timing out)

Added​

  • Windows Environment Support DX-APP now fully supports the Windows operating system! In response to user requests, we've expanded compatibility beyond Linux to include Windows, enabling a broader range of development environments to take advantage of DX-APP.
    : OS: Windows 10 / 11
    : Deepx M1 Driver Version: v1.7.1 or higher
    : Deepx M1 Runtime Lib Version: v3.1.0 or higher
    : Python: Version 3.8 or higher (required for Python module support)
    : Compiler: Visual Studio Community 2022 (required for building C++ examples)
  • Add automated build script (build.bat) for automatic build and Visual Studio solution generation
  • Three new PPU data types : BBOX (for object detection) / POSE (for pose estimation keypoints) / FACE (for face detection landmarks)
  • Enhanced post-processing functions to support PPU inference output format

Known Issues​

  • DeepLabV3 Semantic Segmentation model accuracy may be slightly degraded in dx-compiler(dx_com) v2.1.0. This will be fixed in the next release. The DeepLabV3 model used in the demo was converted using dx-compiler v2.0.0.
  • When using the PPU model for face detection & pose estimation, dx-compiler v2.1.0 does not currently support converting face and pose models to PPU format. This feature will be added in a future release. The PPU models used in the demo were converted using dx-compiler v1.0.0(dx_com v1.60.1).

v2.0.0 / 2025-08-14​

Changed​

  • Moved the YOLO post-processing guide from 07_Python_Examples.md to a new, dedicated document 08_YOLO_Post_Processing_Pybind11.md.
  • Refactored yolo_pybind_example.py to use a RunAsync() + Wait() structure instead of callbacks. This ensures the output tensor order is correctly handled.
  • Major code refactoring and restructuring of demo applications
  • Consolidated common utilities into directory
  • Removed deprecated and legacy codes
  • Update documentation and resources
  • YoloPostProcess now filters and selects the correct tensor by output_name when USE_ORT=ON
  • Command-line help messages in various demos have been improved to clearly mark required parameters.
  • Replaced YOLOv5s-1 example json to YOLOv5s-6 json configuration file has been added for object detection.
  • Documentation has been updated to add Python requirements and modified some images.
  • feat: add OS and architecture checks in build script & update CPU specifications in documentation

Fixed​

  • FPS calculation bug in yolo_multi
  • Removed postprocessing code for legacy PPU models
  • Fixed postprocessing logic to support new output shapes of YOLO models when USE_ORT=OFF
  • fix typo error in framebuffer info file path (yolo_multi app)
  • Improve error messages for output tensor size mismatch and missing in Yolo post processing
  • Rename output tensors in json config 'yolov5s6_example.json'

Added​

  • Added to cleanly purge the pip-installed package and local build artifacts (shared library, dist-info/egg-info, and build directory).
  • Added version guards in templates/python/yolo_pybind_example.py to ensure compatibility with DX-RT ≥ 3.0.0 and DXNN model version ≥ 7
  • Enhanced the JSON configuration to support a target_output_tensor_name key and a name field for each layer parameter.
  • Added a feature to filter output tensor using the target_output_tensor_name provided in the JSON configuration
  • Added a feature to automatically reorder model layer parameters in the JSON configuration to match the model's actual output tensor sequence.
  • Enhanced demo applications
  • Added for easy demo execution
  • Added SCRFD decoding method for run_detector example
  • Added postprocessing support for yolo_pose and yolo_face models (available only when USE_ORT=ON)
  • Listed supported YOLO model types for YoloPostProcess in README
  • feat: add uninstall script and enhance color utility functions

v1.11.0 / 2025-07-24​

Changed​

  • feat: enhance --clean option in build script for pybind artifacts
  • feat: update dxnn models version(1.40.2 to 1.60.1)
  • feat: auto run setup script or display a guide message when a file not found error occurs during example execution

Fixed​

  • feat: Improve error message readability in install, build scripts
    : Apply color to error messages
    : Reorder message output to display errors before help messages
  • Update tensor index assignment in Yolo layer reordering
  • fix: resolve dx_postprocess Python lib build error and improve error handling

v1.10.0 / 2025-06-17​

  • Initial create dx-app
  • demo : classification
  • demo : object detection
  • demo : pose estimation
  • demo : multi models for object detection and segmentation
  • demo : semantic segmentation
  • demo : multi channel oject detection
  • template : classification
  • template : object detection
  • template : python example (sync/async/pybind c++)
Last updated: