shoppy-mobilesam

MobileSAM ONNX Models

MobileSAM models in ONNX format for on-device mobile inference.

Files

• mobile_sam_encoder.onnx + .onnx.data - Image encoder (27MB)
• mobile_sam_decoder.onnx + .onnx.data - Mask decoder (23MB)

Usage

These models are designed for React Native applications using onnxruntime-react-native.

Model URLs

const ENCODER_MODEL_URL = 'https://huggingface.co/gifty-so/shoppy-mobilesam/resolve/main/mobile_sam_encoder.onnx';
const DECODER_MODEL_URL = 'https://huggingface.co/gifty-so/shoppy-mobilesam/resolve/main/mobile_sam_decoder.onnx';

Input Format

Encoder:

• Input: input (float32, shape: [1, 3, 1024, 1024])
  • ImageNet normalized RGB image
• Output: image_embeddings (float32, shape: [1, 256, 64, 64])

Decoder:

• Inputs:
  • image_embeddings (from encoder)
  • point_coords (float32, shape: [1, N, 2]) - Point coordinates normalized to [0, 1024]
  • point_labels (float32, shape: [1, N]) - 1 for positive, 0 for negative
  • mask_input (float32, shape: [1, 1, 256, 256]) - Previous mask or zeros
  • has_mask_input (float32, shape: [1]) - 0 or 1
  • orig_im_size (float32, shape: [2]) - Original image dimensions
• Outputs:
  • masks (float32) - Segmentation masks
  • iou_predictions (float32) - Confidence scores

License

Apache 2.0

Citation

MobileSAM: https://github.com/ChaoningZhang/MobileSAM