XYZ9843/GOOSE-M2F

GOOSE-M2F: Adapting Mask2Former for High-Fidelity, Long-Tailed Fine-Grained Semantic Segmentation in Unstructured Outdoor Terrain

Jyothiraditya Lingam, Nikhileswara Rao Sulake, Sai Manikanta Eswar Machara

Department of Computer Science and Engineering Rajiv Gandhi University of Knowledge Technologies (RGUKT), Nuzvid, Andhra Pradesh, India

📄 Paper • 💻 Code • 🤗 Hugging Face • 🏆 Challenge

GOOSE-M2F is a task-specific adaptation of Mask2Former for the GOOSE 2D Fine-Grained Semantic Segmentation Challenge (ICRA 2026). The proposed framework addresses long-tailed semantic segmentation in unstructured outdoor environments through enhanced object query capacity, feature refinement, auxiliary supervision, class-balanced optimization, and robust multi-scale inference.

Official Challenge Performance: 70.08% Composite mIoU (63.55% Fine mIoU, 76.61% Coarse mIoU), achieving 3rd Place on the GOOSE 2D FGSS Challenge Leaderboard.

---

📢 News

• [ICRA 2026] GOOSE-M2F achieved 3rd Place in the GOOSE 2D Fine-Grained Semantic Segmentation Challenge.
• [2026] Source code and trained models released.
• [2026] Technical report available on arXiv.

---

What is GOOSE-M2F?

The GOOSE dataset presents one of the most challenging real-world segmentation benchmarks: 64 fine-grained classes across diverse unstructured outdoor environments including forests, gravel paths, construction zones, and agricultural terrain — with a severely long-tailed class distribution.

GOOSE-M2F extends the baseline Mask2Former (Swin-Large backbone) with three key modifications engineered specifically for this challenge:

Modification	Problem Solved	Impact
200 Object Queries (vs 100)	Query saturation in 64-class scenes	+2-3% composite mIoU
Feature Refinement Module (FRM) — ASPP-lite + CBAM	Over-segmentation of amorphous terrain classes	+3-4% on Vegetation/Terrain
Auxiliary Supervision Head at H/4 resolution	Vanishing gradients for tiny/thin classes	+5-8% on rare classes

---

Architecture

Input Image [B, 3, H, W]
      │
      ▼
Swin-Large Backbone (Hierarchical, 4 stages)
  Stage 1-4: channels {192, 384, 768, 1536}, resolutions {H/4 → H/32}
      │
      ▼
MSDeformAttn Pixel Decoder (6-layer FPN)
  Output: mask_features [B, 256, H/4, W/4]
      │
      ├──────────────────────────────────────┐
      ▼                                      ▼
[NEW] Feature Refinement Module        [NEW] Auxiliary Head
  ASPP-lite: dilations {1, 3, 6, 12}     Conv(256→256→64)
  + Global Average Pooling               DB-weighted CE loss
  + CBAM Dual-Attention (Ch + Sp)        Supervised at H/4
      │
      ▼
Transformer Decoder (9 layers)
  [MOD] 200 Object Queries (was 100)
  Masked Cross-Attention
      │
      ▼
Class Head [B, 200, 65] × Mask Head [B, 200, H/4, W/4]
      │
      ▼
Hungarian Matching → Semantic Prediction

---

Training Strategy

Technique	Description
Distribution-Balanced (DB) Loss	w_c = (1-β)/(1-β^n_c), β=0.9999. Amplifies gradients for rare classes.
Rare-Class Copy-Paste (RCCP)	Pre-extracted rare-class cutouts pasted onto training images at 85% probability.
Dynamic IoU-Aware Weights	Per-class loss weights updated every epoch from validation IoU (0%→4x, 80%+→1x).
10x LR Jump (V4)	Backbone 1e-5, Decoder 5e-5 — broke the model out of a local minimum at ~55%.
EMA (decay=0.9995)	Shadow weights consistently +1.0–1.5% over raw model on validation.
Class-Aware Repeat Sampling	Oversamples images containing rare classes proportional to their rarity.
Polynomial LR Decay	Gradual decay after warmup, with annealing in final sessions.

Training Progression (V1 → V8)

Session	Base LR	Backbone LR	Official Score
V1 (S3)	5e-6	1e-6	50.68%
V2 (S4)	5e-6	1e-6	54.62%
V3 (S5)	5e-6	1e-6	55.64%
V4 (S6)	5e-5	1e-5	56.38% ← 10x LR Jump
V5 (S7)	5e-5	1e-5	57.59%
V6 (S8)	5e-5	1e-5	58.58%
V7 (S9)	5e-5	1e-5	59.23%
V8 (S10)	2.5e-5	5e-6	59.51% ← Annealing
Inference	—	—	70.08% ← +10.57% from TTA

---

Inference Engine

The final performance leap from 59.51% (training) to 70.08% (submission) came entirely from the inference pipeline:

Technique	Gain	Description
Dense Sliding Window	+4-5%	896×896 crops, stride=384px (57% overlap)
2D Gaussian Kernel Blending	Eliminates artifacts	Center pixels weighted higher, edges down-weighted
4-Scale TTA	+3-4%	Scales: 0.5×, 0.75×, 1.0×, 1.5×
H-Flip TTA	+1-2%	8 total views per image (4 scales × 2 flips)
EMA Weights	+1-1.5%	Shadow weights used instead of raw training weights
AuxHead Stripping	VRAM savings	Removed before inference — not needed for prediction

---

Project Structure

goose-m2f/
├── src/
│   ├── model.py          ← GOOSEMask2Former (FRM + AuxHead + 200 queries)
│   ├── features.py       ← Dataset, augmentations, EMA, metrics
│   ├── train.py          ← Training engine (Trainer class)
│   └── inference.py      ← Dense Gaussian patch-blending inference
├── configs/
│   ├── train_config.yaml ← All training hyperparameters
│   └── infer_config.yaml ← TTA and inference settings
├── data/raw/             ← Dataset (symlink or copy)
├── models/               ← Manually placed checkpoints
├── outputs/
│   ├── checkpoints/      ← best_model.pth, latest.pth, charts
│   └── predictions/      ← Output PNG predictions
├── tests/
│   └── test_model.py     ← pytest unit tests
├── instructions/
│   └── instructions.md   ← Full setup + usage guide
└── requirements.txt

---

Quick Start

1. Setup

git clone https://github.com/Aditya-Lingam-9000/GOOSE-2D-FGSS-Challenge
cd GOOSE-2D-FGSS-Challenge

conda create -n goose python=3.11 -y && conda activate goose
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt
accelerate config   # Configure for your GPU setup

2. Configure Paths

Edit configs/train_config.yaml:

data_dir: "/path/to/goose_dataset"
csv_path: "/path/to/goose_label_mapping.csv"
output_dir: "outputs/checkpoints/session_01"

3. Train

# Single GPU
python -m src.train --config configs/train_config.yaml

# Multi-GPU
accelerate launch --num_processes 2 -m src.train --config configs/train_config.yaml

4. Inference

Edit configs/infer_config.yaml with the checkpoint path and image directory, then:

python -m src.inference --config configs/infer_config.yaml

5. Tests

pytest tests/ -v

---

Results

Official Leaderboard Performance (Final Submission)

Metric	Score
Fine mIoU	~68.5%
Coarse mIoU	~71.6%
Official Composite	70.08%

Coarse Category Breakdown

Category	mIoU
Sky	94.6%
Road	91.0%
Vehicle	89.8%
Vegetation	89.8%
Construction	75.5%
Terrain	78.9%
Human	62.8%
Sign	62.4%
Water	33.9%
Object	51.3%
Animal	0.0%

---

Requirements

Package	Version
torch	≥ 2.1.0
transformers	≥ 4.38.0
accelerate	≥ 0.27.0
albumentations	≥ 1.3.1
opencv-python	≥ 4.9.0
numpy	≥ 1.24.0

See requirements.txt for the complete list.

---

Citation

If you use this work, please cite:

@techreport{---,
  title     = {GOOSE-M2F: Adapting Mask2Former for High-Fidelity, Long-Tailed Fine-Grained Semantic Segmentation in Unstructured Outdoor Terrain},
  author    = {---},
  year      = {2026},
  institution = {---}
}

---

References

• Mask2Former: Cheng et al., Masked-Attention Mask Transformer for Universal Image Segmentation, CVPR 2022
• Swin Transformer: Liu et al., ICCV 2021
• CBAM: Woo et al., Convolutional Block Attention Module, ECCV 2018
• DeepLab: Chen et al., Rethinking Atrous Convolution, TPAMI 2017

---