How sefakurtipek/fdsst_tracking works
Project type and purpose
• This is a single-executable C++ computer-vision application implementing the Fast Discriminative Scale Space Tracking (fDSST) algorithm.
File structure
• Top-level C++ source and header files:
• main.cpp – entry point, parses user input (frames, video, webcam) and drives the tracker
• FDSSTTracker.cpp/.hpp – core tracker class managing initialization, feature extraction, model update, and scale estimation
• fhog_fDSST.* – HOG feature extraction module
• ffttools_fDSST.hpp – FFT routines for fast correlation and scale space computation
• labdata_fDSST.hpp – data structures for loading ground-truth or labelling data
• SSE2NEON.h, sse.hpp – SIMD abstraction layer supporting both x86 SSE2 and ARM NEON optimizations
• Visual Studio solution files (.sln, .vcxproj) indicate a Windows build environment.
• dataset/ holds test sequences and ground-truth files.
README for additional insights
• Confirms algorithm purpose: fDSST tracking in C++
• Mentions that main.cpp was extended to handle different input sources
• Instructs the user to adjust file paths for input data
a. Main components
• Input Manager – handles frame sources (image folder, video file, webcam)
• Tracker Engine – encapsulates FDSSTTracker class
– Correlation & Scale Estimation (uses ffttools_fDSST)
– Model Update (online adaptation of target appearance)
• Utilities
– Geometry Utils (recttools, iou, centerPointDistance)
– Data Structures (labdata_fDSST)
– SIMD Layer (SSE2NEON, sse.hpp)
• Dataset Loader – reads dataset/groundtruth.txt and frame lists for benchmarking
b. Relationships and interactions
• The Input Manager feeds raw frames into the Tracker Engine
• Tracker Engine calls the SIMD Layer during feature extraction for performance
• Feature Extraction and FFT modules interact tightly within the Tracker Engine
• Geometry Utils are invoked by the Tracker Engine to compute overlaps, distances
• Output Module subscribes to Tracker Engine events and writes results or overlays
c. Architectural patterns and principles
• Modular decomposition: separate files for each responsibility (features, FFT, tracking logic, utilities)
• Data flow pipeline: Input → Preprocessing → Feature Extraction → Tracking Estimation → Model Update → Output
• Abstraction of hardware optimizations via SIMD layer for portability and maintainability
d. Key technologies and libraries
• C++11 or later (for standard library containers and algorithms)
• FFT routines (in-house ffttools_fDSST)
• HOG descriptor implementation (fhog_fDSST)
• SIMD abstraction (SSE2NEON for cross-platform vectorization)
• Visual Studio build system (.sln, .vcxproj)