cs766project

Presentation & Evaluation Code Release

This package contains source code for qualitative comparison, depth ablation visualization, attention overlays, and scenario mIoU plots used in our course project website and report.

1. What is included

Component	Description
`presentation_common.py`	Shared loaders, Cityscapes coloring, confusion / IoU
`project_paths.py`	`VIREO_REPO_ROOT` resolution (no hard-coded machine paths)
`qualitative_multimodel_grid.py`	1×5 grid vs FC-CLIP, SED, our model
`depth_zero_ablation.py`	Zero-depth branch ablation
`extract_cross_attention_map.py`	Hook-based heatmap overlay
`scenario_miou_eval.py`	Seen / Unseen mIoU + bar chart
`detectron2_batch_infer.py`	FC-CLIP / SED inference helpers
`run_all_presentations.py`	Orchestrates all figures
`download_baselines.py`	Optional helper to fetch public baseline URLs (run yourself)
`setup_baseline_env.sh`	Example conda recipe for Detectron2 baselines
`assets/`	Cityscapes open-vocab label JSON for SED
`third_party/`	Vendored pointers / clones / README — see each `README`
`docs/DATA_AND_CHECKPOINTS.md`	Policy: no dataset / no checkpoints in zip

2. What is NOT included (deliberate exclusions)

We follow standard academic practice: code release without hosting copyrighted datasets or multi‑GB weights.

Item	Policy
Datasets	Not shipped. Obtain ACDC, Cityscapes, etc. from official sites; arrange paths to match Vireo configs.
Checkpoints	Not shipped. Train the model from the training repo or download authors’ public weights yourself.
Baseline weights	FC-CLIP / SED: use official Google Drive links referenced in upstream repos / `download_baselines.py`.

Teaching staff: Full end-to-end reproduction requires GPU clusters, matching package versions, registered datasets, and trained or downloaded weights—we document steps but do not guarantee one-click reruns.

3. Layout expectations

Recommended monorepo layout (adjust VIREO_REPO_ROOT if different):

your_workspace/
  vireo_repro/
    Vireo/                    # training repo: configs, vireo/, data/, work_dirs/, ...
  DGSS/
    cv_project/               # this package (presentation scripts)

Export:

export VIREO_REPO_ROOT=/absolute/path/to/Vireo
export PYTHONPATH="${VIREO_REPO_ROOT}:${PWD}/DGSS/cv_project:${PYTHONPATH}"

Run scripts with cwd = Vireo when configs use relative paths (e.g. open_vocab/cityscapes.json).

4. Quick sanity check (no data required)

python -c "from pathlib import Path; import sys; sys.path.insert(0,'.'); from project_paths import get_vireo_root; print(get_vireo_root())"

5. Building the downloadable zip

From this repo:

bash website_submission/scripts/pack_release.sh

Produces cv_project_source_release.zip next to DGSS/ by default (custom path: first argument).
The zip excludes .git, checkpoints, .npy caches, compiled .so, and large baseline binaries—while keeping all .py / .sh / docs so the archive looks complete.

6. License

Presentation scripts: Apache-2.0 or project license as stated in the course repo (match your LICENSE file).
Third-party code (FC-CLIP, SED, MMSeg) remains under their original licenses—do not remove upstream license headers.

7. Citation

Cite the original FC-CLIP, SED, MMSegmentation, and your Vireo technical report as required by the course.