Project number 1

• Led a project involving open-source dataset collection, data preprocessing, and training semantic segmentation models (UNet, Swin-Transformer, DeeplabV3)
• Conducted channel-level compression on the UNet network and tested models of various sizes.
• Converted the model format from PyTorch to ONNX to TIDL.
• Deployed the model on the TDA4 platform, performed INT 8 quantization compression, and conducted performance testing with models of different sizes.
• Future work involves collecting a semantic segmentation dataset using a UGV platform for underground parking HD mapping (Multi-sensor fusion).

”