DeepSignal-S1-1.0
GitHub · Source Code & Reproduction
A multimodal System One model for traffic signals: visual observations in, fast structured control decisions out.
English | 简体中文
Model · Scenes · Results · Reproduction
DeepSignal-S1 follows a Jev-style structured probabilistic decision approach. It reads temporal intersection views and signal context, scores candidate actions, and selects a legal control action directly. It does not generate a traffic-analysis essay for another program to parse, and does not require a separate YOLO detection stage.
DeepSignal-S1-1.0 is trained on simulation scenes from nine Nansha intersections, including T-junctions and four-arm intersections, and five traffic-demand scenarios at one four-arm intersection in CARLA's official Town10HD map, covering balanced, uneven and time-varying flows. Its development addresses two practical questions: can visual control improve on the fixed-time plan in clean conditions, and can it remain more robust than detection plus Max Pressure when the camera feed degrades?
Why System One
System One emphasizes fast, directly callable decisions, as illustrated by Jev and the open-source OneJev. DeepSignal-S1 applies this decision paradigm to traffic control: its inference interface returns an action and probabilities without generating a token-by-token reasoning narrative. It is not a wrapper around the commercial Jev service.
- Vision to control: temporal observations and signal context determine actions without operator-supplied live vehicle counts.
- Structured decisions: the model chooses among legal actions; the signal executor implements the committed green tail, yellow and all-red transitions.
- One checkpoint across scenes: the release covers T-junctions and four-arm intersections in Nansha as well as Town10HD.
- Local deployment: unquantized consumer-GPU inference is measured; quantization and edge deployment remain future work.
- Effective clean control: versus the fixed-time plan, S1 reduces vehicle-weighted mean journey time by 18.17% across 72 Nansha cases and 33.77% across five Town regimes.
- Competitive with YOLO+MP in clean views: Nansha mean journey time is a further 1.36% lower, while Town10HD is essentially tied, without a separate detector.
- Robust under visual disruption: fixed camera shift reduces Town mean journey time by 82.93% versus YOLO+MP; 31 px motion blur reduces Nansha mean journey time by 30.14%.
- Unquantized local inference: 6823 RTX 5090 model calls average 485 ms, with a 479 ms median.
Five Town10HD Traffic-Demand Scenarios
All five regimes use the same four-way junction, varying directional demand, load and peak timing rather than road geometry. The table gives the settings used in the current evaluation.
| Demand scenario | Flow per approach (vehicles/hour) | Temporal pattern | Control challenge |
|---|---|---|---|
| Balanced flow | All four: 360 each | Constant throughout | Basic control with similar demand in all directions |
| Sustained directional imbalance | E and W: 600 each N and S: 120 each | Constant major/minor ratio 5:1 | Prioritize persistently busier approaches |
| Higher-load directional imbalance | E and W: 750 each N and S: 150 each | Constant still 5:1 | 25% more total demand than the preceding regime |
| Time-varying peak | All approaches: 200 → 700 → 200 | 0-60 s 60-120 s from 120 s | Adapt to rising demand and peak recovery |
| Shifted peak timing | All approaches: 200 → 200 → 700 | 0-60 s 60-120 s from 120 s | Adapt to changes in peak timing |
Flows are expected arrival rates per approach, not per lane or exact hourly vehicle counts; arrivals are sampled randomly. All regimes use 70% straight, 15% left and 15% right turns. Time starts at simulation startup, with the final rate held until demand generation ends. Shifted peak timing reorders the time-varying rates by seed; the table shows the configuration behind the current results. Controllers receive identical vehicle plans.
Results
Results compare mean waiting time per vehicle, mean queue length and mean journey time per vehicle; lower is better for all three. Clean views compare S1, the fixed-time plan and YOLO+MP. Perturbed views compare S1 with YOLO+MP.
Clean views: three controllers
| Scene | Condition | Cases | Controller | Mean wait (s/veh) | Mean queue (veh) | Mean journey (s/veh) |
|---|---|---|---|---|---|---|
| Nansha | Clean | 72 | S1 | 45.82 | 10.47 | 107.32 |
| Fixed-time plan | 68.55 | 15.52 | 131.15 | |||
| YOLO+MP | 46.89 | 10.56 | 108.79 | |||
| Town10HD | Clean | 5 | S1 | 20.45 | 3.56 | 38.22 |
| Fixed-time plan | 40.46 | 5.40 | 57.70 | |||
| YOLO+MP | 20.22 | 3.59 | 38.22 |
Perturbed views: S1 vs YOLO+MP
| Scene | Condition | Cases | Controller | Mean wait (s/veh) | Mean queue (veh) | Mean journey (s/veh) |
|---|---|---|---|---|---|---|
| Nansha | Defocus σ=9 | 9 | S1 | 76.08 | 15.79 | 141.03 |
| YOLO+MP | 104.86 | 31.04 | 167.48 | |||
| Nansha | 31 px motion blur | 9 | S1 | 67.38 | 17.23 | 131.57 |
| YOLO+MP | 124.59 | 30.18 | 188.34 | |||
| Nansha | Fixed yaw ±15° / pitch ±8° | 9 | S1 | 63.68 | 15.44 | 129.04 |
| YOLO+MP | 87.01 | 21.18 | 154.36 | |||
| Nansha | Dynamic shake 5° | 9 | S1 | 63.40 | 16.17 | 128.25 |
| YOLO+MP | 72.15 | 17.91 | 137.61 | |||
| Town10HD | Defocus σ=9 | 5 | S1 | 27.51 | 4.10 | 44.43 |
| YOLO+MP | 92.97 | 9.75 | 109.15 | |||
| Town10HD | 31 px motion blur | 5 | S1 | 20.28 | 3.58 | 38.10 |
| YOLO+MP | 111.26 | 10.24 | 127.71 | |||
| Town10HD | Fixed yaw ±15° / pitch ±8° | 5 | S1 | 19.23 | 3.45 | 37.11 |
| YOLO+MP | 205.09 | 13.67 | 217.40 | |||
| Town10HD | Dynamic shake 5° | 5 | S1 | 19.13 | 3.32 | 37.20 |
| YOLO+MP | 21.49 | 3.53 | 39.88 |
Metrics: lower is better for all three. Mean wait includes stopped and insertion waiting; mean queue counts halted vehicles; journey time starts at planned departure, with unfinished vehicles contributing elapsed time at the evaluation cutoff. Detailed definitions.
What the Results Show
Clean Views: Competitive Control Without a Separate Detector
Against the fixed-time plan, S1 reduces mean journey time by 18.17% in Nansha and 33.77% in Town10HD, while mean waiting falls by 33.15% and 49.45%, respectively. The gains extend beyond earlier arrivals to less stopped waiting and shorter queues.
Against YOLO+MP, S1 further reduces Nansha mean journey time by 1.36% and mean waiting by 2.28%; Town10HD mean journey time is essentially tied. Direct visual decisions therefore deliver competitive traffic control without a separate vehicle-detection stage, with a modest clean-view lead in Nansha.
Visual Disruption: A More Distinctive Deployment Advantage
Traffic demand may stay the same while camera input changes because of lighting, maintenance or vibration. The four tests target distinct imaging problems: can S1 continue making effective signal decisions when images lose sharpness or stability, or no longer match the original viewpoint?
- Defocus: inaccurate focus softens vehicle outlines. Mixed lighting, including headlights at night, can disrupt autofocus; manual focus adjustment or lens replacement can also leave the intersection out of focus. Softer vehicle and lane edges make recognition and counting harder. Under simulated defocus, S1 mean journey time is 15.79% lower than YOLO+MP in Nansha and 59.29% lower in Town.
- Motion blur: movement during exposure creates directional smearing. In low light at dusk or night, a camera may use longer exposures. Moving vehicles or camera movement during exposure can then smear outlines and obscure separation between vehicles. Under the motion-blur test, S1 mean journey time is 30.14% lower in Nansha and 70.16% lower in Town.
- Fixed camera-angle shifts: a stable view from a changed direction. For example, maintenance-related re-aiming or a shifted mount can move lane and stop-line positions in the image, even after the camera becomes steady. Previously drawn lane regions may no longer align. Our comparison calibrates YOLO at the original pose, then holds the changed camera angle fixed without recalibration. S1 mean journey time is 16.40% lower in Nansha and 82.93% lower in Town; Town mean waiting is 90.63% lower.
- Dynamic shake: the viewpoint keeps moving over time. Wind gusts or vibrations from nearby heavy vehicles can shake a pole-mounted camera, shifting vehicle and lane positions across successive frames. Unlike a fixed angle shift, this disturbance continues over time. S1 retains a mean-journey advantage of 6.80% in Nansha and 6.72% in Town. Shake still yields gains, while blur and camera shifts show the more pronounced advantages.
These changes can affect vehicle detection and counting or misalign the image with calibrated lane regions, with downstream consequences for count-based signal control. S1 decides directly from temporal images and signal state, without that separate detection-and-counting pipeline. Across both evaluated scenes and all four perturbations, S1 has lower mean waiting, queue length and journey time than YOLO+MP: robustness translates into less waiting, shorter queues and shorter average journeys.
The tests simulate visual changes that these real-world problems can produce; they are not completed night-time field or real-world adverse-weather validation. Real-world context and technical sources.
Sub-Second Decisions From a Visual Model
Across 6823 unquantized RTX 5090 decision requests, latency averages 485 ms, with 99.08% below one second; steady-state requests average 477 ms. System One emits structured control actions directly, without waiting for a long textual response, combining visual understanding, traffic-control performance and fast decisions.
Evaluation scale: clean views cover 72 cases across nine Nansha intersections and five Town10HD demand regimes. Each of the four perturbations covers nine Nansha cases and five Town10HD cases, using one traffic seed per scene.
Lane regions are calibrated on the clean camera pose and then frozen. For angle15, the camera is shifted after calibration, remains fixed for the episode, and is not recalibrated. shake5 applies dynamic yaw/pitch peaks of 5° and roll peaks of 2.5°.
Inference Performance
| Scope | Count | Mean | Median | P95 | Max | Under 1 s |
|---|---|---|---|---|---|---|
| All calls | 6823 | 485.4 ms | 479.0 ms | 494.0 ms | 2304.0 ms | 99.08% |
| First call per run | 133 | 911.1 ms | 495.0 ms | 1460.2 ms | 2304.0 ms | 52.63% |
| Subsequent calls | 6690 | 476.9 ms | 479.0 ms | 493.0 ms | 675.0 ms | 100.00% |
Statistics cover 133 S1 evaluation cases and 6823 decision requests with the unquantized BF16 model. Latency measures model decision requests, excluding camera capture, rendering and signal execution; first calls include initialization effects.
Deployment value: sub-second model decisions make consumer-GPU signal-control workstations a practical target. We also tested the unquantized model on an RTX 4090, observing approximately 900 ms per decision. This independent observation uses measurement conditions that have not yet been aligned and is not included in the table. Quantization and edge deployment are future directions for reducing compute requirements.
Download and Quickstart
Model and simulation-asset downloads require separate access approvals. Demo GIFs, videos and result data are openly available from DeepSignal-S1-Media. The release retains the v3 scene assets; use them with the latest GitHub code. Viewing media does not require downloading the scene package.
python3.12 -m venv .venv-inference
source .venv-inference/bin/activate
pip install -U huggingface_hub
hf auth login
hf download AIMS2025/DeepSignal-S1 --local-dir ./DeepSignal-S1-model
deactivate
git clone https://github.com/AIMS-Laboratory/DeepSignal-S1.git DeepSignal-S1
cd DeepSignal-S1
python3.10 -m venv .venv
.venv/bin/pip install -e '.[test,reproduction,download]'
The model server, a real three-frame request that needs no simulator, and closed-loop commands are documented in Reproduction. Download the complete model folder; approved access does not create a hosted API. A full fresh-download closed-loop replay has not yet been certified.
Evaluation Scope
- Sole fixed baseline: the fixed-time plan, with timing parameters held constant throughout evaluation; Nansha uses the published per-intersection plans and Town10HD uses
31/6/31/6 s. - Visual comparison: clean views compare all three controllers; perturbations compare S1 with YOLO+MP. Camera shifts occur after YOLO lane-region calibration, without recalibration.
- Scope: results come from controlled Nansha and Town10HD simulations; deployment on real roads and under real weather remains to be validated.
See evaluation protocol, metric definitions, and media delivery.
Historical Releases
Earlier DeepSignal-S1-1.0-candidate Nansha results and the separate historical Town checkpoint remain archived with their original protocol labels, distinct from the current 1.0 release and its results. See historical released results.
Contact: dalei@scu.edu.cn
简体中文
面向交通信号控制的多模态 System One（端到端快思考）模型：从路口视觉直接作出快速、结构化的控制决策。
English | 简体中文
DeepSignal-S1 采用 Jev 式结构化概率决策路线：读取连续路口画面和信号状态，直接对候选控制动作给出概率并选择合法动作，而不是先生成长篇交通分析，再由外部程序解析文字。视觉理解与动作选择在同一个模型中完成，不依赖额外的YOLO车辆检测环节。
DeepSignal-S1-1.0 的训练覆盖南沙9个路口的仿真场景（包括丁字与十字交叉口），以及 CARLA 官方 Town10HD 地图中1个十字交叉口的5类交通需求场景，涵盖均衡、不均衡和时变车流。项目围绕两个实际问题展开：正常视觉下，能否比固定配时方案更有效；摄像头失焦、运动模糊或机位变化后，能否比“目标检测＋最大压力控制”更稳健。
为什么是 System One
System One（端到端快思考）强调快速、直接、可被软件调用的决策。Jev与开源OneJev体现了这一模型范式。DeepSignal-S1 将这种决策方式用于交通信号控制，当前推理接口直接返回动作及其概率，不逐字生成解释或推理过程；它不是调用商业Jev服务。
- 从视觉到控制： 时序路口画面与当前信号上下文共同决定动作，无需人工提供实时车辆计数。
- 结构化输出： 在当前合法动作中直接决策，信号执行器负责既定的绿灯尾段、黄灯和全红转换。
- 适应不同路口： 同一模型覆盖南沙的丁字及十字路口和Town10HD场景，不为每个场景更换权重。
- 面向本地部署： 使用未量化模型验证消费级GPU上的决策延迟，量化与边缘部署仍是后续工作。
- 正常控制有效： 相对固定配时方案，平均行程时间在南沙72案例降低 **18.17%**，在Town五类车流降低 **33.77%**。
- 正常视觉对标YOLO+MP： 南沙平均行程进一步降低 **1.36%**，Town10HD基本持平，无需独立车辆检测环节。
- 特殊视觉稳健： 固定机位偏移下，Town平均行程比YOLO+MP降低 **82.93%**；31像素运动模糊下，南沙降低 **30.14%**。
- 未量化本地推理： RTX 5090共6823次模型请求，平均 485毫秒，中位数479毫秒。
Town10HD的5类交通需求
五类车流均设置在同一个十字路口，区别在方向分配、流量大小和高峰出现时点，而不是五个不同路口。下表列出当前评测使用的设置。
| 交通需求场景 | 每个进口方向的流量 （辆/小时） | 时间变化 | 考察重点 |
|---|---|---|---|
| 均衡车流 | 东、西、南、北各360 | 全程恒定 | 四向需求相近时 的基本控制能力 |
| 持续方向不均衡 | 东、西各600 南、北各120 | 全程恒定 主次方向流量比5:1 | 持续向高需求方向 分配绿灯 |
| 更高流量的不均衡 | 东、西各750 南、北各150 | 全程恒定 仍为5:1 | 比上一类总需求增加25% 考察较高负荷 |
| 时变高峰 | 四向均为 200 → 700 → 200 | 0-60秒 60-120秒 120秒起 | 适应车流上升 及高峰消退 |
| 高峰时序变化 | 四向均为 200 → 200 → 700 | 0-60秒 60-120秒 120秒起 | 改变高峰出现时点 考察时序适应能力 |
流量是每个进口方向的期望到达率，不是每车道流量，也不是每小时固定生成的车辆数；车辆按该流率随机到达。五类均采用直行70%、左转15%、右转15%。时间从仿真启动计，末档流量保持至需求生成结束。“高峰时序变化”按种子重排时变流量，上表为当前结果对应的配置。各控制器使用相同的车辆计划。
评测结果
以下比较平均每车等待时间、平均排队长度和平均每车行程时间，三项指标均越低越好。正常视觉比较S1、固定配时方案和YOLO+MP三方；特殊视觉比较S1与YOLO+MP。
正常视觉：三方比较
| 场景 | 条件 | 案例数 | 控制方法 | 平均等待（秒/车） | 平均排队长度（辆） | 平均行程（秒/车） |
|---|---|---|---|---|---|---|
| 南沙 | 正常 | 72 | S1 | 45.82 | 10.47 | 107.32 |
| 固定配时方案 | 68.55 | 15.52 | 131.15 | |||
| YOLO+MP | 46.89 | 10.56 | 108.79 | |||
| Town10HD | 正常 | 5 | S1 | 20.45 | 3.56 | 38.22 |
| 固定配时方案 | 40.46 | 5.40 | 57.70 | |||
| YOLO+MP | 20.22 | 3.59 | 38.22 |
特殊视觉：S1与YOLO+MP
| 场景 | 条件 | 案例数 | 控制方法 | 平均等待（秒/车） | 平均排队长度（辆） | 平均行程（秒/车） |
|---|---|---|---|---|---|---|
| 南沙 | 失焦σ=9 | 9 | S1 | 76.08 | 15.79 | 141.03 |
| YOLO+MP | 104.86 | 31.04 | 167.48 | |||
| 南沙 | 31像素运动模糊 | 9 | S1 | 67.38 | 17.23 | 131.57 |
| YOLO+MP | 124.59 | 30.18 | 188.34 | |||
| 南沙 | 固定偏航±15° / 俯仰±8° | 9 | S1 | 63.68 | 15.44 | 129.04 |
| YOLO+MP | 87.01 | 21.18 | 154.36 | |||
| 南沙 | 动态抖动峰值5° | 9 | S1 | 63.40 | 16.17 | 128.25 |
| YOLO+MP | 72.15 | 17.91 | 137.61 | |||
| Town10HD | 失焦σ=9 | 5 | S1 | 27.51 | 4.10 | 44.43 |
| YOLO+MP | 92.97 | 9.75 | 109.15 | |||
| Town10HD | 31像素运动模糊 | 5 | S1 | 20.28 | 3.58 | 38.10 |
| YOLO+MP | 111.26 | 10.24 | 127.71 | |||
| Town10HD | 固定偏航±15° / 俯仰±8° | 5 | S1 | 19.23 | 3.45 | 37.11 |
| YOLO+MP | 205.09 | 13.67 | 217.40 | |||
| Town10HD | 动态抖动峰值5° | 5 | S1 | 19.13 | 3.32 | 37.20 |
| YOLO+MP | 21.49 | 3.53 | 39.88 |
指标说明： 三项指标均越低越好。平均等待包括停车等待和入网等待；平均排队长度以停车车辆数计；平均行程从计划出发计时，未完成车辆计入评估截止时的已耗时。详细指标定义。
结果解读
正常视觉：高水平控制，不依赖独立车辆检测
与固定配时方案相比，S1在南沙和Town10HD的平均行程分别降低 18.17% 和 **33.77%**，平均等待分别降低 33.15% 和 **49.45%**。收益不仅是车辆更早到达，也体现在更少的停车等待与更短的排队。
与YOLO+MP相比，S1在南沙的平均行程再降低 **1.36%**、平均等待降低 **2.28%**；Town10HD的平均行程基本持平。这说明，直接从视觉作出信号决策，可以在省去独立检测环节的同时达到有竞争力的交通控制效果，并在南沙取得小幅领先。
特殊视觉：更值得关注的部署优势
现实路口的车流未必改变，摄像头输入却可能因光照变化、设备维护或风振而发生变化。四类测试分别对应不同的成像问题，考察S1在画面不再清晰、稳定或保持原机位时，能否继续作出有效的信号决策。
- 失焦：对焦不准，车辆轮廓变软。 夜间车灯等复杂光源可能干扰自动对焦，人工调焦或更换镜头后也可能没有重新调好焦点。车辆边缘和车道细节随之变得模糊，增加识别与计数的难度。在失焦测试中，S1相对YOLO+MP的平均行程在南沙、Town分别降低 **15.79%、59.29%**。
- 运动模糊：曝光期间的移动形成拖影。 傍晚或夜间光照不足时，摄像头可能采用更长曝光；车辆运动或曝光期间的相机晃动便可能形成方向性拖影，使车辆形状和相邻车辆的边界更难分辨。在运动模糊测试中，S1的平均行程分别降低 **30.14%、70.16%**。
- 固定角度偏移：调整后稳定在新的机位。 例如检修后重新调整视角，或支架发生偏转：画面已经稳定，但停车线和车道在图像中的位置改变了，原先画好的车道区域可能不再对应。我们的对比先在正常机位为YOLO画线，再固定改变相机角度，不重新画线。S1的平均行程在南沙降低 **16.40%**，在Town降低 **82.93%**；Town的平均等待同时降低 **90.63%**。
- 动态抖动：视角随时间来回变化。 高杆摄像头可能受到阵风或邻近重型车辆引起的振动影响，车辆和车道在连续画面中的位置随之摆动；这与调整后保持不动的固定偏角不同。动态抖动测试中，S1的平均行程在南沙、Town分别降低 **6.80%、6.72%**。抖动下仍有收益，模糊和机位偏移下的优势更突出。
这些变化既可能影响车辆检测与计数，也可能使画面与原有车道区域错位，进而影响依赖检测计数的信号控制。S1直接依据时序画面和信号状态决策，不依赖这条独立检测计数链路。在本次两个场景、四种特殊视觉条件的对比中，S1的平均等待、平均排队长度和平均行程均低于YOLO+MP，体现出对这些输入变化的鲁棒性，优势直接落在车辆少等待、少排队和平均行程更短上。
上述测试模拟的是这些现实问题可能造成的视觉变化，不等同于已经完成夜间现场或真实恶劣天气验证。现实场景对应与资料依据。
亚秒级决策：视觉模型也能快速响应
未量化模型在RTX 5090上的 6823 次决策请求平均耗时 485毫秒，99.08% 的请求低于1秒；稳态请求平均 477毫秒。System One直接输出结构化控制动作，无需等待长篇文字生成，让视觉理解、交通控制效果与快速响应能够同时兼顾。
评测规模： 正常视觉覆盖南沙9个路口的72案例和Town10HD的5类车流；四种特殊视觉条件下，每种分别覆盖南沙9案例和Town10HD 5案例。特殊视觉为单种子仿真评测。
YOLO车道画线先在正常机位完成并冻结。angle15在画线后改变机位，整回合保持固定，不重新标定；shake5的偏航和俯仰峰值为5°，横滚峰值为2.5°。
推理性能
| 范围 | 数量 | 平均 | 中位数 | P95 | 最大值 | 小于1秒 |
|---|---|---|---|---|---|---|
| 全部请求 | 6823 | 485.4 ms | 479.0 ms | 494.0 ms | 2304.0 ms | 99.08% |
| 每回合首请求 | 133 | 911.1 ms | 495.0 ms | 1460.2 ms | 2304.0 ms | 52.63% |
| 后续请求 | 6690 | 476.9 ms | 479.0 ms | 493.0 ms | 675.0 ms | 100.00% |
统计覆盖133个S1评测案例、6823次决策请求，使用未量化BF16模型。延迟仅计算模型决策请求，不含相机采集、渲染及信号执行；首请求包含初始化影响。
部署价值： 亚秒级模型决策让消费级GPU信控工作站成为可考虑的落地方向。同时，我们也在 RTX 4090 上测试了未量化模型的推理速度，单次决策约 900毫秒；这是一项独立测试观测，测量条件尚未对齐，不计入上表。后续将探索量化与边缘部署，进一步降低算力需求。
下载与快速开始
模型与仿真资产下载分别需要申请审批。GIF、视频和结果数据在DeepSignal-S1-Media免费公开，无需审批。当前版本沿用v3场景资产并配合GitHub最新代码；观看媒体无需下载场景包。
python3.12 -m venv .venv-inference
source .venv-inference/bin/activate
pip install -U huggingface_hub
hf auth login
hf download AIMS2025/DeepSignal-S1 --local-dir ./DeepSignal-S1-model
deactivate
git clone https://github.com/AIMS-Laboratory/DeepSignal-S1.git DeepSignal-S1
cd DeepSignal-S1
python3.10 -m venv .venv
.venv/bin/pip install -e '.[test,reproduction,download]'
模型服务、无需启动仿真器的真实三帧请求示例，以及闭环命令见复现说明。请下载完整模型目录；模型获批不代表提供远程托管API。从公开下载开始的完整闭环复跑尚未完成认证。
评测说明
- 唯一固定基线： 固定配时方案，评测期间保持配时参数不变；南沙使用已发布的各路口配时表，Town10HD使用
31/6/31/6秒。 - 视觉对照： 正常视觉比较三方，特殊视觉比较S1与YOLO+MP；相机偏移发生在YOLO画线之后，不重新标定。
- 适用范围： 结果来自南沙与Town10HD的受控仿真，真实道路和天气条件下的部署效果仍需验证。
历史版本
此前 DeepSignal-S1-1.0-candidate 的南沙结果和历史Town专用权重继续按原协议归档，与当前1.0版本及其成绩区分。见历史发布结果。
- Downloads last month
- 415