How aiclub-uit/aicity2026-track2 works
Team 24’s product is an AI City Challenge Track 2 reproduction pipeline: it prepares public-test inputs, trains or loads vision-language adapters, generates VQA answers and traffic-scene captions, then applies temporal decoding and caption fact corrections before writing submissions. The README documents the overall stages and optional Cosmos-based bundle retraining. Source excerpts show test preparation invoking preprocessing, context construction using WSD helpers, and Cosmos preparation reusing VQA frame-selection logic; unsampled training and pipeline details are represented only at the documented subsystem level.
Main entry points: run_qwen35_vqa.py VQA adapter · run_e2e.py E2E orchestrator · preprocess_vqa.py VQA preprocessing · run_qwen35_caption.py Caption adapters
Components 19
Data preparation
- Challenge datasets
data - Test layout
code/prepare_test_root.py - VQA preprocessing
code/preprocess_vqa.py - Caption preprocessing
code/preprocess_caption.py - Context builder
code/ctx_builder.py
Adapters and inference
- VQA adapter
code/run_qwen35_vqa.py - Caption adapters
code/run_qwen35_caption.py - Shipped LoRA adapters
artifacts/adapters - Cosmos preparation
code/cosmos_restyle_prep.py - Cosmos restyling
- Domain randomization
code/dr_prep.py
Reasoning and refinement
- Temporal decoding
code/wsd.py - Fact correction v2
code/fact_stitch.py - Fact correction v3
code/fact_stitch3.py - Fact correction v4
code/fact_stitch4.py - Fact correction v5
code/fact_stitch5.py
Pipeline and outputs
- E2E orchestrator
code/run_e2e.py - Submission files
submissions
Other components
- Reproducer
Connections 27
- Reproducer → E2E orchestrator: runs stages
README.md - Challenge datasets → Test layout: supplies test data
README.md - Test layout → VQA preprocessing: invokes preprocessing
- Test layout → Caption preprocessing: invokes preprocessing
- Challenge datasets → VQA preprocessing: supplies frames and annotations
README.md - Challenge datasets → Caption preprocessing: supplies frames and annotations
README.md - VQA preprocessing → Context builder: provides prepared evidence
- VQA preprocessing → VQA adapter: provides prediction inputs
README.md - Caption preprocessing → Caption adapters: provides prediction inputs
README.md - Context builder → Temporal decoding: uses question helpers
code/ctx_builder.py - Context builder → VQA adapter: supports context route
- Shipped LoRA adapters → VQA adapter: loads adapter
README.md - Shipped LoRA adapters → Caption adapters: loads adapter
README.md - Cosmos preparation → VQA preprocessing: reuses frame selection
code/cosmos_restyle_prep.py - Cosmos preparation → Cosmos restyling: prepares restyle inputs
README.md - Cosmos restyling → Domain randomization: provides restyled frames
README.md - Domain randomization → VQA adapter: supports bundle training
README.md - VQA adapter → Temporal decoding: provides probabilities
README.md - Temporal decoding → E2E orchestrator: decodes answers
code/run_e2e.py - Caption adapters → Fact correction v2: provides draft captions
- Fact correction v2 → Fact correction v3: refines captions
code/fact_stitch3.py - Fact correction v3 → Fact correction v4: refines captions
code/fact_stitch4.py - Fact correction v4 → Fact correction v5: refines captions
code/fact_stitch5.py - E2E orchestrator → VQA adapter: trains and predicts
README.md - E2E orchestrator → Caption adapters: trains and predicts
README.md - E2E orchestrator → Submission files: writes results
README.md - Fact correction v5 → Submission files: writes corrected output
code/fact_stitch5.py