How melantech/gola works
This repository implements GOLA, a parameter-efficient RGB-T visual-tracking research system. The main workflow prepares pretrained DINOv2 weights with SVD-based rank grouping, launches configurable training or evaluation through a shared runtime, consumes tracking datasets, and produces checkpoints, metrics, and tracking results for benchmark evaluation. Training-pair sampling and image augmentation, plus evaluation task scheduling and result collection, are distinct data stages. The graph combines documented capabilities with sampled call references; unsampled internals are represented at their existing directory paths, and their uncertain links are omitted.
Main entry points: builder.py Runtime application · main.py CLI bootstrap · worker.py Training data worker · default Training runner
Components 22
Startup and runtime
- CLI bootstrap
trackit/core/boot/main.py - Config loading
trackit/core/boot/funcs/main/load_config.py - Global context
trackit/core/boot/funcs/main/init_global_context.py - Application setup
trackit/core/boot/funcs/main/init_application.py - Runtime application
trackit/core/runtime/application/default/builder.py - Checkpoint service
trackit/core/runtime/services/checkpoint.py
Model and training
- Rank grouping
pretrained_models/params_svd_grouping.py - GOLA tracker
trackit/models/methods/GOLA/gola.py - DINOv2 backbone
trackit/models/backbone/dinov2/__init__.py - Training criteria
trackit/criteria/builder.py - Training runner
trackit/runner/training/default
Tracking data
- Pair sampling
trackit/data/methods/siamese_tracker_train/siamese_training_pair_sampling/worker.py - Training data worker
trackit/data/methods/siamese_tracker_train/worker.py - Data augmentation
trackit/data/methods/siamese_tracker_train/transform/default/augmentation/pipeline.py - Tracking datasets
trackit/data/source/dataset/MMOT.py
Evaluation and results
- Evaluation data worker
trackit/data/methods/siamese_tracker_eval/worker.py - Task scheduler
trackit/data/sampling/per_frame/distributed_dynamic_eval_task_scheduling/distributed_evaluation_task_scheduler.py - Evaluation orchestration
trackit/data/methods/siamese_tracker_eval/orchestration.py - Result collection
trackit/data/components/result_collector/collector.py - Benchmark metrics
trackit/data/components/result_collector/handler/one_pass_evaluation/ope_metrics.py
Other components
- Researcher
- RGB-T benchmarks
Connections 24
- Researcher → CLI bootstrap: launches
- CLI bootstrap → Config loading: loads config
trackit/core/boot/main.py - CLI bootstrap → Global context: initializes context
trackit/core/boot/main.py - CLI bootstrap → Application setup: prepares app
trackit/core/boot/main.py - Application setup → Runtime application: builds runtime
trackit/core/boot/funcs/main/init_application.py - Researcher → Rank grouping: prepares weights
- Rank grouping → GOLA tracker: groups model ranks
- Runtime application → GOLA tracker: runs tracker
- GOLA tracker → DINOv2 backbone: uses features
- Runtime application → Training runner: runs training
- Training runner → Training criteria: optimizes with
- Training runner → Training data worker: requests batches
- Training data worker → Pair sampling: samples pairs
trackit/data/methods/siamese_tracker_train/worker.py - Training data worker → Data augmentation: transforms samples
- Training data worker → Tracking datasets: reads frames
trackit/data/methods/siamese_tracker_train/worker.py - Runtime application → Evaluation orchestration: runs evaluation
- Evaluation orchestration → Evaluation data worker: requests inputs
trackit/data/methods/siamese_tracker_eval/orchestration.py - Evaluation data worker → Task scheduler: gets tasks
trackit/data/methods/siamese_tracker_eval/worker.py - Evaluation data worker → Tracking datasets: reads sequences
trackit/data/methods/siamese_tracker_eval/worker.py - Evaluation orchestration → Result collection: collects results
- Result collection → Benchmark metrics: computes metrics
- Runtime application → Checkpoint service: saves weights
- Result collection → RGB-T benchmarks: exports tracking results
- RGB-T benchmarks → Researcher: reports scores