brittleru/nser-ibvs-mask-splitter-dataset

NSER-IBVS Mask Splitter Dataset

Dataset Description

This dataset is used to train the Mask Splitter neural network, a key component of the NSER-IBVS visual servoing framework for autonomous drone control. The network learns to split a vehicle segmentation mask into front and back regions, enabling the analytical IBVS controller to compute precise velocity commands.

Associated Resources

Resource	Link
Paper	ICCV 2025 Workshop
arXiv	2507.19878
Models	nser-ibvs-models
Annotation and Train	mask-splitter
NN Integration for drone	nser-ibvs-drone
Project	Website
Demo	Hugging Face Space

Sample Data

RGB Image

Segmentation

Front Mask

Back Mask

Intended Use

This dataset is intended for:

• Training and evaluating mask-splitting or part-aware segmentation models
• Visual servoing and robotics perception research
• Simulation-to-real transfer studies

Out of scope:

• Generic object detection benchmarks
• Autonomous driving datasets

Dataset Structure

data/
├── sim/                           # Simulation data (UE4 Bunker environment)
│   ├── train/
│   │   ├── images/                # RGB frames
│   │   │   ├── around-car-30-45-60-75-90-high-quality/
│   │   │   ├── around-car-30-45-60-75-90-low-quality/
│   │   │   ├── around-car-90-75-60-45-30-low-quality/
│   │   │   ├── just-environment-high-quality/
│   │   │   └── just-environment-low-quality/
│   │   ├── labels/                # Manually annotated front/back masks with mask-splitter code.
│   │   │   └── <scene>/
│   │   │       ├── front/
│   │   │       └── back/
│   │   └── segmented/             # Full segmentation masks
│   └── validation/
│       ├── images/
│       ├── labels/
│       └── segmented/
└── real/                          # Real-world data
    ├── train/
    └── validation/

Data Format

Component	Format	Resolution	Description
images/	PNG	640x360	RGB frames from drone camera
segmented/	PNG (binary)	640x360	Full vehicle mask
labels/front/	PNG (binary)	640x360	Manually annotated front vehicle region
labels/back/	PNG (binary)	640x360	Manually annotated back vehicle region

Naming Convention: Files share the same name across images/, segmented/, and labels/ directories for simple correspondence (e.g., frame_000000_1076195.png).

Scenes

Simulation (UE4 Bunker Environment)

Training scenes:

• around-car-30-45-60-75-90-high-quality - Various angles, high render quality
• around-car-30-45-60-75-90-low-quality - Various angles, low render quality
• around-car-90-75-60-45-30-low-quality - Reverse angle sequence
• just-environment-high-quality - Environment-only frames (negatives)
• just-environment-low-quality - Environment-only frames (negatives)

Validation scenes:

• around-car-45-high-quality
• around-car-45-low-quality
• around-car-45-low-quality-car-at-45

Real-World

Captured with Parrot Anafi 4K drone tracking a real vehicle.

Training scenes:

• real-30-45-60-75-90 - Various angles
• just-environment-real - Environment-only frames (negatives)

Validation scenes:

• real-val

Usage

With git

git lfs install
git clone https://huggingface.co/datasets/brittleru/nser-ibvs-mask-splitter-dataset

Direct File Access

from huggingface_hub import snapshot_download

# Download entire dataset
snapshot_download(
    repo_id="brittleru/nser-ibvs-mask-splitter-dataset",
    repo_type="dataset",
    local_dir="./mask-splitter-data"
)

# Or download specific subset
snapshot_download(
    repo_id="brittleru/nser-ibvs-mask-splitter-dataset",
    repo_type="dataset",
    local_dir="./mask-splitter-data-sim",
    allow_patterns="data/sim/*"
)

Load with Hugging Face Datasets

from datasets import load_dataset

# Load specific config and split
ds_sim_train = load_dataset("brittleru/nser-ibvs-mask-splitter-dataset", "sim", split="train")
ds_sim_val = load_dataset("brittleru/nser-ibvs-mask-splitter-dataset", "sim", split="validation")
ds_real_train = load_dataset("brittleru/nser-ibvs-mask-splitter-dataset", "real", split="train")

example = ds_sim_val[0]
print(example["scene"])        # Scene name
example["image"].show()        # RGB image (PIL)
example["front_mask"].show()   # Front mask (PIL)
example["back_mask"].show()    # Back mask (PIL)

PyTorch DataLoader

import torch
from datasets import load_dataset
from torch.utils.data import DataLoader
from torchvision import transforms

ds = load_dataset("brittleru/nser-ibvs-mask-splitter-dataset", "sim", split="train")

transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
])

def collate_fn(batch):
    images = []
    targets = []
    
    for x in batch:
        # RGB image (3 channels)
        img = transform(x["image"].convert("RGB"))
        # Segmentation mask (1 channel)
        seg_mask = transform(x["segmentation_mask"].convert("L"))
        # Concatenate to get 4-channel input
        input_4ch = torch.cat([img, seg_mask], dim=0)
        images.append(input_4ch)
        
        # Stack front and back masks as target (2 channels)
        front = transform(x["front_mask"].convert("L"))
        back = transform(x["back_mask"].convert("L"))
        target = torch.cat([front, back], dim=0)
        targets.append(target)
    
    return torch.stack(images), torch.stack(targets)

dataloader = DataLoader(ds, batch_size=16, shuffle=True, collate_fn=collate_fn, num_workers=4)

Training the Mask Splitter

# Clone the main repository
git clone https://github.com/SpaceTime-Vision-Robotics-Laboratory/mask-splitter.git

# Install requirements

# Run train script (check mask-splitter README.md for additional arguments)
python runnable/train_splitter_network.py --data_dir=/path/to/your/data

Inference Example

import cv2
from mask_splitter.nn.infer import MaskSplitterInference

splitter = MaskSplitterInference(
    model_path="path/to/mask_splitter.pt",
    device="cuda"
)

image = cv2.imread("frame.png")
mask = cv2.imread("segmented.png", cv2.IMREAD_GRAYSCALE)

front_mask, back_mask = splitter.infer(image, mask)
splitter.visualize(image, front_mask, back_mask)

Dataset Statistics

Split	Domain	Images	With Vehicle	Environment Only
Train	Sim	14,693	10,114	4,579
Train	Real	13,760	9,118	4,642
Val	Sim	1,123	1,123	-
Val	Real	1,084	1,084	-
Total	-	30,660	21,439	9,221

Simulation Data

Train:

Scene	Images	Type
around-car-30-45-60-75-90-high-quality	2,034	Vehicle
around-car-30-45-60-75-90-low-quality	3,212	Vehicle
around-car-90-75-60-45-30-low-quality	4,868	Vehicle
just-environment-high-quality	2,129	Environment
just-environment-low-quality	2,450	Environment

Subtotal: 14,693 images

Validation:

Scene	Images	Type
around-car-45-high-quality	343	Vehicle
around-car-45-low-quality	391	Vehicle
around-car-45-low-quality-car-at-45	389	Vehicle

Subtotal: 1,123 images

Real-World Data

Train:

Scene	Images	Type
just-environment-real	4,642	Environment
real-30-45-60-75-90	9,118	Vehicle

Subtotal: 13,760 images

Validation:

Scene	Images	Type
real-val	1,084	Vehicle

Subtotal: 1,084 images

Limitations and Biases

• Vehicle category is limited primarily to toy cars.
• Camera viewpoint is drone-mounted (top-down / oblique).
• Lighting conditions are limited by the simulator and indoor real-world captures.
• No nighttime data is included.

Citation

If you use this dataset in your research, please cite:

@InProceedings{Mocanu_2025_ICCV,
    author    = {Mocanu, Sebastian and Nae, Sebastian-Ion and Barbu, Mihai-Eugen and Leordeanu, Marius},
    title     = {Efficient Self-Supervised Neuro-Analytic Visual Servoing for Real-time Quadrotor Control},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2025},
    pages     = {1744-1753}
}