AIMS2025/DeepSignal-S1-Scenes

DeepSignal-S1 Simulation Scenes

One combined simulation bundle for Nansha and Town10HD.

Test Code and Results · Model · Exact Reproduction Protocol · 简体中文

This repository distributes simulation runtimes and test inputs for DeepSignal-S1, a vision-first traffic signal controller. It is a downloadable asset bundle, not a tabular dataset for datasets.load_dataset. No training implementation or model weights are included. Google Drive is not required.

Bundle Contents

File	Size	Release
DeepSignal-S1-Scenes-v2.tar.gz	27.60 GiB (29,632,627,613 bytes)	v2

The single archive includes:

• UE5/CARLA 0.10.0 Nansha runtime and nine-intersection test materials.
• UE5/CARLA 0.10.0 Town10HD runtime for the migrated/supplementary suite.
• The original CARLA 0.9.16 Docker image for the archived Town10HD suite.
• Matching Python 3.10 CARLA client wheels, fixed arrival plans, intersection/camera configurations and engineering fixed-timing plans.
• The stock YOLO11m detector used by the independent YOLO+MP baseline, with its upstream license.
• Test-only reproduction code and component notices/licenses.

Download and Install

Linux and sufficient disk space for both the compressed archive and extracted runtimes are required. NVIDIA GPU rendering, SUMO 1.18.0 and separate Python environments are needed for visual closed-loop tests. Docker with NVIDIA CDI is required only for the archived Town runtime.

pip install -U huggingface_hub
hf download AIMS2025/DeepSignal-S1-Scenes DeepSignal-S1-Scenes-v2.tar.gz \
  --repo-type dataset --local-dir ./downloads

git clone https://github.com/AIMS-Laboratory/DeepSignal-S1.git
cd DeepSignal-S1
python3.10 -m venv .venv
.venv/bin/pip install -e '.[test,reproduction]'
.venv/bin/deepsignal-s1 install-assets \
  --archive ../downloads/DeepSignal-S1-Scenes-v2.tar.gz
.venv/bin/pip install assets/runtime/python/carla-0.10.0-cp310-cp310-linux_x86_64.whl

Use the latest GitHub test code and release-specific commands. The archive is the existing v2 bundle, unchanged; documentation snapshots inside it may contain superseded hosting instructions. This dataset and the current GitHub instructions replace those download instructions.

Town launcher correction (2026-10-26): the archive's code snapshot predates a UE5 camera-pose initialization fix. Use the current GitHub test code with these runtime assets, not the bundled test-code snapshot. New GitHub runs also collect intersection queue, stopped-time and throughput metrics. The scene binaries do not need to be downloaded again for this code correction.

Town timing correction: subsequent checks also found that legacy Town adaptive executors omitted the committed five-second green tail, initialized MP with synthetic zero counts, and could stop one second early. Current GitHub code corrects these issues and records actual signal transitions; all 240 current-checkpoint Town runs are restarted, with earlier records retained and excluded. Historical Town scores are not validated under the corrected protocol. Nansha measurements are unaffected by this Town correction. No scene archive replacement is needed.

Model and Result Scope

The dataset is public; the model repository separately requires approved access. Downloading scenes does not grant model access or provide a hosted API. Use the complete approved model folder for local inference.

The current HF checkpoint, DeepSignal-S1-1.0-candidate, corresponds to completed Nansha measurements. Published historical Town10HD measurements use a different archive checkpoint, which is not included in this dataset or the current HF model repository. The scene bundle supplies its runtime but not its weights. Do not attribute historical Town scores to the current model or expect the runtime alone to reproduce them. Current-checkpoint supplementary measurements are separate from the published tables.

A complete fresh-download closed-loop rerun has not yet been certified. Published measurements apply to the documented versions, cases, metrics and camera conditions, not arbitrary scenes or a bitwise guarantee across machines.

Licenses and Notices

This is a multi-component distribution, not a single blanket license. Preserve the CARLA, Unreal Engine, operating-system, map, material and Ultralytics notices accompanying their respective components. Archived container notices remain inside the image, with the CARLA license also supplied alongside it. Review the third-party notices and the licenses in the archive before use or redistribution. Synthetic traffic plans are not measurements of real-world demand.

简体中文

南沙与 Town10HD 共用这一份场景包，不再需要 Google Drive。 本数据集仅提供仿真运行文件、固定车流及测试材料，不含训练实现或模型权重，也不是供 datasets.load_dataset 读取的表格数据。

文件与内容

DeepSignal-S1-Scenes-v2.tar.gz 大小为 27.60 GiB（29,632,627,613 字节），包含南沙九路口、UE5 Town10HD、历史 Town 所用的 CARLA 0.9.16 Docker 镜像、两种版本匹配的 Python 3.10 wheel、固定车流与配时、相机配置、独立 YOLO+MP 对照所用的原版检测器，以及必要许可。

按上方命令下载并安装。需要 Linux、足够容纳压缩包和解压文件的磁盘空间；视觉闭环需要 NVIDIA GPU、SUMO 1.18.0，以及分别配置的模型推理和仿真客户端环境。历史 Town 环境还需要 Docker 和 NVIDIA CDI。两个 CARLA 版本的 wheel 不要安装到同一客户端环境。

测试请使用 GitHub 最新代码和发行版本复现说明。本次直接托管已有 v2 压缩包，没有重新打包；包内旧文档若仍提其他托管方式，以本页和 GitHub 当前下载说明为准。

2026-10-26 Town 入口修正： 包内测试代码快照早于 UE5 相机位姿初始化修复，请使用本包运行素材配合 GitHub 当前测试代码，不要直接使用旧测试快照。新版代码另记交叉口排队、停车等待与通过量，无需为此重新下载三维运行素材。

Town 执行时序修正： 后续检查还发现旧 Town 自适应执行器漏掉了承诺的 5 秒绿灯尾段、MP 首轮使用人造全零观测，以及部分回合提前一秒结束。GitHub 当前代码已统一修正并记录实际信号变化，当前权重的 Town 全部 240 回合重新运行，旧记录保留并排除。历史 Town 数字不作为修正后协议的已验证成绩，南沙测量不受此次 Town 修正影响；无需更换场景压缩包。

模型与复现范围

场景公开下载，模型仍须到 HF 模型主页申请权限。当前 HF 权重 DeepSignal-S1-1.0-candidate 对应已完成的南沙测量；已发布的历史 Town 成绩来自另一份存档权重，该权重既不包含在本场景包，也不在当前 HF 模型仓库内。提供历史运行环境，不等于已提供历史模型，不能将其成绩归给当前权重。

补测与已发布表格分开。从公开下载开始的完整闭环复跑尚未验收；不承诺任意硬件、软件版本或场景下数值逐位一致。各组件遵循自身许可，不以统一许可证覆盖第三方条款，使用或再分发前请查看包内许可与 第三方说明。

联系 / Contact: dalei@scu.edu.cn