ilessio-aiflowlab/DEF-tuni

TUNI: Real-Time RGB-T Semantic Segmentation

Paper: TUNI: Real-time RGB-T Semantic Segmentation with Unified Multi-Modal Feature Extraction and Cross-Modal Feature Fusion Venue: ICRA 2026 Authors: Xiaodong Guo et al.

Model

• Architecture: TUNI (Conv + Attention hybrid) with unified RGB-T encoder
• Variant: 384_2242 (dims=[48,96,192,384], depths=[2,2,4,2])
• Parameters: 10.6M
• Input: RGB (3×480×640) + Thermal (3×480×640)
• Output: Segmentation mask (480×640)

Performance

Dataset	Classes	mIoU	Pixel Acc
FMB	15	62.4%	91.4%
PST900	5	TBD	TBD
CART	12	TBD	TBD

Available Formats

• tuni_fmb.pth — PyTorch state dict
• tuni_fmb.safetensors — Safetensors format
• tuni_fmb.onnx + tuni_fmb.onnx.data — ONNX (opset 18)
• tuni_fmb_fp16.trt — TensorRT FP16 engine
• tuni_fmb_fp32.trt — TensorRT FP32 engine

Usage

from def_tuni.model import TUNIModel

model = TUNIModel(variant="384_2242", n_classes=15)
model.load_checkpoint("tuni_fmb.pth")
model.eval().cuda()

# Inference
pred = model(rgb_tensor, thermal_tensor)
segmentation = pred.argmax(dim=1)

ANIMA Module

Part of the ANIMA Defense Module ecosystem. Built with ANIMA by Robot Flow Labs.