Paper: MCOD: First Multispectral Camouflaged Object Detection Benchmark (arXiv: 2509.15753)
Architecture: ResNet-50 8-channel backbone + Spectral Attention + PRNet Progressive Decoder
Parameters: 33.3M
Training: Phase 1 RGB pretraining on COD10K (4,040 images)
Input: 8-channel tensor (3 RGB + 4 NIR + 1 thermal), 352x352
Output: Binary camouflage segmentation mask

Training Results (Phase 1)

Metric	Value
val/loss	2.256
val/iou	0.226
Epochs	50
Batch size	44 per GPU
GPUs	2x NVIDIA L4 (DDP)
Precision	FP16 mixed

Available Formats

Format	File	Size
PyTorch (.pth)	`project_ragnarok_cuda_v1_phase1_best.pth`	128MB
SafeTensors	`project_ragnarok_cuda_v1_phase1_best.safetensors`	128MB
ONNX (opset 17)	`project_ragnarok_v1_phase1.onnx`	111MB
TensorRT FP32	`project_ragnarok_v1_phase1_trt_fp32.engine`	114MB
TensorRT FP16	`project_ragnarok_v1_phase1_trt_fp16.engine`	57MB

Usage

python

import torch
from anima_ragnarok.model import RagnarokModel

model = RagnarokModel(config_path="configs/model_config.yaml")
state = torch.load("project_ragnarok_cuda_v1_phase1_best.pth")
model.load_state_dict(state["model_state_dict"], strict=False)
model.eval()

# Input: 8-channel image [B, 8, 352, 352]
output = model(input_tensor)
mask = output[0]  # [B, 1, 352, 352]

ANIMA Module

Part of the ANIMA robotics perception stack. See RobotFlow-Labs/defense-project-ragnarok.

ilessio-aiflowlab/project_ragnarok

RAGNAROK - Multispectral Camouflaged Object Detection

Model Details

Training Results (Phase 1)

Available Formats

Usage

ANIMA Module

ilessio-aiflowlab/
project_ragnarok