ForestSim: A Synthetic Benchmark for Intelligent Vehicle Perception in Unstructured Forest Environments arXiv: 2603.27923 Authors: Pragat Wagle, Zheng Chen, Lantao Liu

Architecture

DeepLabV3 with ResNet-50 backbone, trained on the ForestSim 24-class synthetic off-road segmentation dataset. 39.6M parameters.

Classes (24)

grass, tree, pole, water, sky, vehicle, container, asphalt, gravel, mulch, rockbed, log, bicycle, person, fence, bush, sign, rock, bridge, concrete, table, building, void, generic ground

Results

Metric	Paper (m2)	Reproduced	Delta
mIoU	61.87%	61.42%	-0.45
aAcc	89.85%	89.07%	-0.78
mAcc	—	70.15%	—

Reproduced within 0.5 mIoU on single NVIDIA L4 (23GB) with batch size 24.

Exported Formats

Format	File	Size	Use Case
PyTorch (.pth)	`pytorch/vis_forestsim_v1.pth`	317MB	Training, fine-tuning
SafeTensors	`pytorch/vis_forestsim_v1.safetensors`	159MB	Fast loading, safe
ONNX	`onnx/vis_forestsim_v1.onnx`	159MB	Cross-platform inference
TensorRT FP16	—	—	Generate on target hardware
TensorRT FP32	—	—	Generate on target hardware

CUDA Kernels

Three custom CUDA kernels built for this module (shared at anima-cuda-infra):

Kernel	Speedup	Purpose
`seg_argmax_colorize`	2.6x	Fused argmax + palette colorize
`terrain_roughness`	<0.1ms	Gradient-based traversability scoring
`mask_morphology`	5ms	Erosion/dilation/majority vote cleanup

Usage

python

import torch
import torchvision.models.segmentation as seg

# Load model
model = seg.deeplabv3_resnet50(num_classes=24)
ckpt = torch.load("pytorch/vis_forestsim_v1.pth", map_location="cpu")
model.load_state_dict(ckpt["model"])
model.eval()

# Inference
image = torch.randn(1, 3, 512, 512)
with torch.no_grad():
    logits = model(image)["out"]  # [1, 24, 512, 512]
    mask = logits.argmax(1)       # [1, 512, 512]

Training

Hardware: NVIDIA L4 (23GB VRAM)
Batch size: 24 (auto-detected, 68% VRAM)
Optimizer: SGD (lr=0.01, momentum=0.9, weight_decay=5e-4)
Scheduler: Warmup (500 steps) + Cosine decay
Iterations: 38,500 (early stopped from 40,000)
Precision: FP16 mixed precision
Dataset: ForestSim 24-class (1884 train / 209 test)
Config: See configs/paper_deeplabv3_r50.toml

Docker

bash

docker compose -f docker-compose.serve.yml --profile serve up -d
curl localhost:8080/health

ROS2

Subscribes to /camera/image_raw, publishes:

/vis_forestsim/segmentation — colorized 24-class mask
/vis_forestsim/traversability — 6-class traversability mask

License

Apache 2.0 — Robot Flow Labs / AIFLOW LABS LIMITED

ilessio-aiflowlab/project_vis_forestsim

VIS-FORESTSIM — ANIMA Off-Road Segmentation Module

Paper