BananaMind FaceDetect V1

BananaMind FaceDetect V1 is an early face detection model from Banaxi Tech. It uses a custom Hugging Face Transformers-compatible wrapper around torchvision.models.detection.ssdlite320_mobilenet_v3_large and is intended for research and experimentation.

This model is licensed under CC-BY-NC-ND-4.0 and inherits non-commercial/no-derivatives usage restrictions from the WIDER FACE dataset license listed on Hugging Face. Do not treat this model as Apache, MIT, or commercial-use safe.

Model Details

• Repository: Banaxi-Tech/BananaMind-FaceDetect-V1
• Model name: BananaMind FaceDetect V1
• Task: Face detection / object detection
• Architecture: SSDLite320 MobileNetV3 Large via torchvision
• Framework wrapper: Custom Hugging Face Transformers remote code
• Training: Trained from scratch
• Pretrained backbone weights: None
• Ultralytics/YOLO: Not used
• Number of classes: 2
• Class 0: background
• Class 1: face

The repository includes:

• model.safetensors: main safetensors checkpoint
• original_model.pt: original PyTorch checkpoint
• config.json: custom Transformers config
• configuration_banaxi_face_detector.py: custom config class
• modeling_banaxi_face_detector.py: custom model wrapper
• README.md: model card

Dataset And Attribution

This model was trained on WIDER FACE using the Hugging Face dataset repository CUHK-CSE/wider_face.

WIDER FACE is a face detection benchmark containing face bounding boxes across images with large variation in face scale, pose, occlusion, and scene type. The Hugging Face dataset page lists the dataset license as CC-BY-NC-ND-4.0.

Please also cite and follow the terms for WIDER FACE when using this model:

@inproceedings{yang2016wider,
  title={WIDER FACE: A Face Detection Benchmark},
  author={Yang, Shuo and Luo, Ping and Loy, Chen Change and Tang, Xiaoou},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2016}
}

Training Notes

• Dataset source: CUHK-CSE/wider_face
• Dataset task: face detection benchmark
• Training hardware: NVIDIA A100 80GB GPU
• Training used converted WIDER FACE bounding boxes
• Very tiny faces were filtered out during training because the detector input size is 320px and tiny WIDER FACE boxes destabilized early training
• Base detector: ssdlite320_mobilenet_v3_large
• Input format: RGB image tensor normalized to 0..1

Usage

This repository uses custom remote code:

• configuration_banaxi_face_detector.py
• modeling_banaxi_face_detector.py

Load the model with trust_remote_code=True.

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained(
    "Banaxi-Tech/BananaMind-FaceDetect-V1",
    trust_remote_code=True,
)
model.eval()

# pixel_values must be a tensor shaped B x 3 x H x W.
# Values should be RGB floats in the range 0..1.
pixel_values = torch.rand(1, 3, 320, 320)

with torch.no_grad():
    outputs = model(pixel_values=pixel_values)

prediction = outputs[0]
boxes = prediction["boxes"]    # xyxy pixel coordinates
scores = prediction["scores"]  # confidence scores
labels = prediction["labels"]  # label 1 means face

keep = scores >= 0.5
face_boxes = boxes[keep]
face_scores = scores[keep]

Outputs follow the torchvision detection format:

• boxes: predicted bounding boxes in xyxy pixel coordinates
• scores: confidence scores
• labels: predicted class labels

Label 1 means face. A confidence threshold around 0.5 to 0.7 is recommended, depending on the use case.

Limitations

• This is an early V1 model.
• It was trained from scratch and may be less accurate than large pretrained face detectors.
• It may miss very small faces.
• It may produce duplicate or partial boxes in difficult cases.
• It is intended mainly for experimentation and research.
• The model inherits usage restrictions from the WIDER FACE dataset license.

License

This model is released under CC-BY-NC-ND-4.0.

The training dataset, WIDER FACE as hosted at CUHK-CSE/wider_face, is also listed on Hugging Face as CC-BY-NC-ND-4.0. Because of these terms, this model should be used for research/non-commercial purposes and should not be presented as commercially usable or permissively licensed.