One combined asset bundle supplies Nansha and Town10HD. It contains simulation runtimes and test inputs, not model weights or training implementation, and is not a tabular dataset for datasets.load_dataset.

Sole Fixed Baseline

Frozen engineering timing is the only fixed baseline. Nansha uses the frozen per-intersection engineering tables (J02: 35/14/10/17 seconds); Town uses 31/6/31/6 seconds. No alternative searched or tuned fixed plan is offered.

Clean views: S1, frozen engineering fixed timing and YOLO+MP are compared.
Perturbed views: only S1 and YOLO+MP are compared; fixed scores are neither run nor summarized for those conditions.
The archive manifest records these engineering tables and this comparison policy. Configurations, reference fixed measurements and the GitHub runner use the same plans.
Historical other-fixed references and associated videos are retained in internal archives, not relabeled as engineering timing or provided as current reproduction options.

Bundle and Installation

The current file is DeepSignal-S1-Scenes-v3.tar.gz. This revision replaces the reproduction-code and documentation snapshot; the runtime binaries and model weights are unchanged. Previous archive versions remain historical revisions rather than current download options.

The bundle includes UE5/CARLA 0.10.0 Nansha and Town10HD, the original CARLA 0.9.16 archived Town Docker image, matching Python 3.10 client wheels, frozen traffic plans, camera/intersection configuration, the stock YOLO11m detector and component licenses. No model weights are included.

pip install -U huggingface_hub
hf download AIMS2025/DeepSignal-S1-Scenes DeepSignal-S1-Scenes-v3.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-v3.tar.gz
.venv/bin/pip install assets/runtime/python/carla-0.10.0-cp310-cp310-linux_x86_64.whl

Use Linux, SUMO 1.18.0 and sufficient disk space for the archive and extracted assets. Visual tests require NVIDIA rendering and a separate model inference environment. Docker with NVIDIA CDI is needed for the archived Town runtime. CARLA 0.10.0 and 0.9.16 wheels must be installed in separate client environments.

Checkpoints and Historical Timing

The dataset is public. The model repository separately requires approved access; downloading scenes does not grant model access or a hosted API.

The current published DeepSignal-S1-1.0-candidate checkpoint has Nansha reference measurements. Historical Town measurements used a different archive checkpoint, not included here or in the current model repository. Fused-checkpoint results are also separate and must not be attributed to this published checkpoint.

Historical Town adaptive executors omitted the committed five-second green tail. Those raw measurements remain historical, not validated results under the corrected town-5s-green-tail-v2 protocol. The current reproduction code retains the corrected timing; Nansha measurements are unaffected by that Town correction. See execution scope.

A complete fresh-download visual closed-loop rerun has not yet been certified. This packaging update is not a new GPU evaluation and does not promise arbitrary-scene or bitwise reproducibility.

Licenses

Preserve all CARLA, Unreal Engine, operating-system, map/material and Ultralytics notices and licenses. This distribution does not apply one blanket license to third-party components. Review the archive licenses and third-party notices before redistribution. Synthetic traffic plans are not real-world traffic measurements.

简体中文

当前统一场景包为 **DeepSignal-S1-Scenes-v3.tar.gz**。南沙和Town共用此包，不需要Google Drive。v3只更新测试代码与说明快照，不改变三维运行文件，不包含模型权重或训练实现。

唯一固定对照是已冻结的工程固定配时：南沙各路口工程表（J02为35/14/10/17秒），Town为31/6/31/6秒。正常视觉三方比较S1、工程固定配时与YOLO+MP；特殊视觉只比较S1与YOLO+MP，不运行或汇总固定方案。不再提供其他搜索调优固定表作为当前复现选项。旧结果和旧视频内部归档，不能改名冒充工程方案。

包内清单记录工程表及三方/双方比较规则，配置、已完成固定记录与GitHub复现入口使用相同方案。正常视觉下S1相对YOLO+MP最多慢5%，每种极端视觉条件下整体至少快10%是后续开发目标，不是已验证结论，也不要求每个交叉口都获胜。补测先做小批检查，再决定扩量。

按照上方命令下载和安装。需要Linux、SUMO 1.18.0、足够磁盘空间及对应Python环境；视觉测试另需NVIDIA GPU与模型权限，历史Town运行环境另需Docker和NVIDIA CDI。两个CARLA版本的Python wheel不能装到同一个环境。

当前HF发布权重、历史Town专用权重与融合权重的成绩分别标注，不相互继承。历史Town执行器漏掉5秒绿灯尾段的记录不作为修正后协议的新成绩；当前包中的测试代码保留修正后的执行时序，南沙不受该Town修正影响。

从公开下载开始的完整视觉闭环复跑尚未验收；这次更换复现快照不代表新增GPU实验。旧压缩包只保留为历史版本，不作为当前下载入口。各组件遵循自身许可，使用和再分发前须查看包内许可及第三方说明。

联系 / Contact: dalei@scu.edu.cn

Downloads last month: 35

Total file size:

29.6 GB