lllllllama__rigorpilot-skills/examples/annotated-readme-demo/after-setup/ANNOTATED_README.md

📄 README · RigorPilot annotations

🟢 success · evaluation · trusted · SUMMARY · COMMANDS · LOG · status.json

Section coverage: 🟢 1 · 🟣 1 · 🔵 2 · ⚪ 10 (14 sections) · score 0.571

_{🟢 success · 🔵 not executed · ⚪ read only · 🟡 partial / assets missing · 🔴 blocked · 🟣 decision needed — original content unchanged; its relative links resolve against the repo root.}

---

_{⚪ Read only}

MiniSeg: Simple and Efficient Semantic Segmentation

Paper | Project Page | Colab

MiniSeg is a lightweight encoder-decoder for semantic segmentation. With a hierarchical local-global mixer and a two-layer MLP head, MiniSeg-B0 reaches 41.2 mIoU on ADE20K with only 3.8M parameters, running at 118 FPS on a single V100.

_{⚪ Read only}

Highlights

• Simple: no attention approximations, no custom CUDA kernels — pure PyTorch.
• Efficient: 3.8M–13.1M parameters, 8.4G–15.9G FLOPs at 512×512.
• Strong: competitive with models 5–10× larger on ADE20K and Cityscapes.
• Reproducible: all configs, logs, and checkpoints released.

_{⚪ Read only}

News

• [2026-06-12] MiniSeg-B0 / B1 checkpoints and full training logs released.
• [2026-05-30] MiniSeg is accepted to NeurIPS 2026.
• [2026-04-18] Preprint released on arXiv.

_{⚪ Read only}

Model Zoo

ADE20K

Model	Crop	Iters	mIoU (SS)	mIoU (MS)	#Params	FLOPs	Checkpoint	Log
MiniSeg-B0	512×512	160k	41.2	42.0	3.8M	8.4G	miniseg_b0.pth	log
MiniSeg-B1	512×512	160k	44.7	45.5	13.1M	15.9G	miniseg_b1.pth	log

Cityscapes

Model	Crop	Iters	mIoU (SS)	#Params	Checkpoint
MiniSeg-B1	1024×1024	160k	79.9	13.1M	miniseg_b1_city.pth

_{⚪ Read only}

Installation

Tested with Python 3.10, PyTorch 2.3, CUDA 12.1.

conda create -n miniseg python=3.10 -y
conda activate miniseg
pip install torch==2.3.0 torchvision==0.18.0 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

Note

🔵 Folded into the setup plan · not executed directly conda create -n miniseg python=3.10 -y conda activate miniseg pip install torch==2.3.0 torchvision==0.18.0 --index-url https://download.pytorch.org/whl/cu121 … +1 _{Evidence: SUMMARY · COMMANDS · LOG · status.json · tier: code-development}

Data Preparation

Download ADE20K and Cityscapes, then link them under data/:

python tools/prepare_ade20k.py --root data/ade20k
python tools/prepare_cityscapes.py --root data/cityscapes

Expected layout:

data/
├── ade20k/
│   ├── images/{training,validation}
│   └── annotations/{training,validation}
└── cityscapes/
    ├── leftImg8bit/{train,val}
    └── gtFine/{train,val}

Note

🔵 Folded into the setup plan · not executed directly python tools/prepare_ade20k.py --root data/ade20k python tools/prepare_cityscapes.py --root data/cityscapes _{Evidence: SUMMARY · COMMANDS · LOG · status.json · tier: code-development}

Evaluation

Download a checkpoint from the Model Zoo into checkpoints/, then evaluate MiniSeg-B0 on the ADE20K validation split (single-scale):

python tools/eval.py --config configs/miniseg_b0_ade20k.yaml --checkpoint checkpoints/miniseg_b0.pth

Multi-scale + flip evaluation:

python tools/eval.py --config configs/miniseg_b0_ade20k.yaml --checkpoint checkpoints/miniseg_b0.pth --ms --flip

Tip

🟢 Executed successfully（low risk） Command: python tools/eval.py --config configs/miniseg_b0_ade20k.yaml --checkpoint checkpoints/miniseg_b0.pth Observed metrics: mIoU=41.18 · aAcc=79.85 _{Evidence: SUMMARY · COMMANDS · LOG · status.json · tier: result-match}

Training

Train MiniSeg-B0 on ADE20K with 8 GPUs (160k iterations, syncBN):

torchrun --nproc_per_node=8 tools/train.py --config configs/miniseg_b0_ade20k.yaml

Single-GPU debugging run:

python tools/train.py --config configs/miniseg_b0_ade20k.yaml --debug

Important

🟣 Training not run · requires explicit authorization (high-impact) torchrun --nproc_per_node=8 tools/train.py --config configs/miniseg_b0_ade20k.yaml python tools/train.py --config configs/miniseg_b0_ade20k.yaml --debug The trusted lane never launches training on its own; it starts with startup verification only after you approve. _{Evidence: SUMMARY · COMMANDS · LOG · status.json · tier: code-development}

Project Structure

miniseg/
├── miniseg/
│   ├── models/       # backbone, mixer blocks, MLP head
│   └── datasets/     # ADE20K / Cityscapes loaders and transforms
├── tools/            # train / eval / data preparation entrypoints
├── configs/          # experiment configs
└── checkpoints/      # place downloaded checkpoints here

_{⚪ Read only}

FAQ

Q: Evaluation numbers differ slightly from the paper? A: Make sure you use single-scale (--ms off) and torch 2.3; cuDNN kernels changed between 2.x releases and can shift mIoU by ±0.1.

Q: Do you support Windows? A: Training is Linux-only; evaluation works on Windows with the same commands.

_{⚪ Read only}

Citation

@inproceedings{miniseg2026,
  title     = {MiniSeg: Simple and Efficient Semantic Segmentation},
  author    = {Lin, Jia and Ito, Sora and Novak, Petra},
  booktitle = {NeurIPS},
  year      = {2026}
}

_{⚪ Read only}

Acknowledgements

Built on top of mmsegmentation and timm. We thank the ADE20K and Cityscapes teams for the datasets.

_{⚪ Read only}

License

This project is released under the MIT License.

_{⚪ Read only}