VI-sem PRJ. Goal: build a panoptic segmentation model for the IDD 20k dataset (India Driving Dataset) and submit to the AutoNUE '19 leaderboard.
In panoptic segmentation, every pixel gets a class label, and every instance of a thing class (vehicles + living things, level3Ids 4–12) gets its own ID.
FinalTry/
├── data/ # Dataset (gitignored)
│ ├── IDD_Segmentation/ # Part I (you placed this)
│ ├── idd20kII/ # Part II (you placed this)
│ └── idd_full/ # Merged root for createLabels.py (generated)
├── external/ # Reference repos from AutoNUE (gitignored)
│ ├── public-code/ # Ground-truth prep (createLabels.py)
│ └── panopticapi/ # Official evaluation
├── src/
│ ├── datasets/ # IDD panoptic dataset loader (TODO)
│ └── models/ # Model architecture (model TBD)
├── scripts/
│ ├── merge_idd.py # Merge Part I + Part II into idd_full/
│ ├── prepare_panoptic.sh # Run createLabels.py to generate panoptic GT
│ └── evaluate.sh # Run panopticapi evaluation on predictions
├── configs/ # Per-experiment YAML configs
├── outputs/ # Checkpoints + predictions (gitignored)
├── notebooks/ # Exploration
├── requirements.txt
└── README.md
-
Python environment (conda, Python 3.10). System Python is 3.14, too new for PyTorch / Detectron2 / MMDet wheels — so we use a conda env on 3.10 that will host both data prep and the eventual model training.
conda create -n autonue python=3.10 -y conda activate autonue pip install -r requirements.txt
-
External repos. Already cloned to
external/. To re-clone:git clone https://github.com/AutoNUE/public-code.git external/public-code git clone https://github.com/AutoNUE/panopticapi.git external/panopticapi python scripts/patch_external.py # apply local compat patches (idempotent)The patch script handles modern-Pillow compatibility and Windows multiprocessing in
public-code. See EXPERIMENTS.md for what each patch does and why. -
Dataset. See data/README.md for the merge + prep steps.
[IDD Part I + Part II]
│
▼ scripts/merge_idd.py
[data/idd_full/]
│
▼ scripts/prepare_panoptic.sh
[gtFine/{train,val}_panoptic/* + *.json] ← official panoptic ground truth
│
▼ (model — TBD)
[predictions/]
│
▼ scripts/evaluate.sh
[PQ / SQ / RQ metrics]
- Kirillov et al. Panoptic Segmentation. CVPR 2019. paper
- AutoNUE / IDD: http://idd.insaan.iiit.ac.in/
- Public code: https://github.com/AutoNUE/public-code
- Panoptic eval: https://github.com/AutoNUE/panopticapi
- COCO panoptic format: https://research.mapillary.com/eccv18/#panoptic