YOLOX-Pylon-S

A YOLOX-S detector extended with one additional class, traffic cones (traffic_cone), on top of the 80 COCO classes, for 81 classes total. We train the adaptation so the original COCO capabilities are kept, not traded away. In practice the retained COCO score lands above the official YOLOX-S baseline, while the added cone class scores higher than every one of the 80 original classes.

Cones are our public demo class. The same adaptation recipe adds arbitrary custom classes such as defects, parts, or PPE to a proven detector without losing what it already knows.

Part of the YOLOX-Pylon family, S · M · L · XL. This is the fastest checkpoint in the family.

Built by Empirisch Tech GmbH (Vienna, Austria) under our Chaperone AI brand. See About Empirisch Tech below.

Results

We evaluate on COCO val2017 plus a held-out traffic-cone split, 81 classes in a single pass, 640×640 input, IoU 0.50:0.95 unless noted.

Metric	Value
mAP 50:95 (81 classes)	42.0
mAP 50:95, original 80 COCO classes only	41.6
AP50 / AP75	60.2 / 45.8
AP small / medium / large	24.5 / 46.1 / 53.5
AR@100	56.3
traffic_cone AP / AR	74.5 / 77.7
Inference (forward + NMS, batch 1, A100)	1.72 ms

Two things stand out.

• Retention is better than free. The official YOLOX-S val2017 baseline is 40.5 mAP on COCO. After adding the cone class, this checkpoint scores 41.6 on the same 80 classes, which is 1.1 points above the base we started from. We attribute that to the additional training schedule.
• The added class is the best class. At 74.5 AP, traffic_cone outscores all 80 original classes on this checkpoint. The next best are stop sign (71.8), giraffe (69.5), and bear (69.3).

Full per-class AP (81 classes)

class	AP	class	AP	class	AP
person	55.632	bicycle	30.285	car	51.554
motorcycle	44.826	airplane	65.935	bus	69.032
train	61.938	truck	46.615	boat	26.122
traffic light	38.914	fire hydrant	69.017	stop sign	71.756
parking meter	47.091	bench	29.711	bird	35.062
cat	62.883	dog	58.504	horse	59.151
sheep	49.890	cow	54.719	elephant	65.400
bear	69.281	zebra	66.671	giraffe	69.549
backpack	17.542	umbrella	39.401	handbag	13.586
tie	30.176	suitcase	41.589	frisbee	62.697
skis	24.562	snowboard	32.899	sports ball	46.214
kite	44.614	baseball bat	28.172	baseball glove	38.296
skateboard	50.527	surfboard	36.032	tennis racket	45.279
bottle	38.133	wine glass	33.070	cup	39.283
fork	33.245	knife	16.844	spoon	15.319
bowl	42.512	banana	25.952	apple	18.219
sandwich	30.984	orange	29.727	broccoli	22.596
carrot	23.408	hot dog	32.177	pizza	52.505
donut	47.748	cake	39.168	chair	32.042
couch	44.568	potted plant	27.697	bed	42.693
dining table	30.536	toilet	65.092	tv	56.373
laptop	60.517	mouse	59.313	remote	23.742
keyboard	50.499	cell phone	33.708	microwave	53.325
oven	33.296	toaster	39.092	sink	39.782
refrigerator	54.511	book	13.306	clock	49.399
vase	36.430	scissors	28.823	teddy bear	44.074
hair drier	1.065	toothbrush	18.352	traffic_cone	74.516

Comparison with the base model

The comparison that matters is against the checkpoint we adapted from, with the same architecture, the same parameter count, the same FLOPs, and one extra class. Baseline figures are the official COCO val2017 numbers from the YOLOX model table.

Model	COCO mAP 50:95	Params	FLOPs	Custom classes	License
yolox_pylon_s (this model)	41.6 kept + traffic_cone 74.5	9.0M	26.8G	cone added, COCO kept	Apache-2.0
YOLOX-S (base)	40.5	9.0M	26.8G	COCO only	Apache-2.0

Reading that table, the adaptation costs us nothing on the original 80 classes. The retained score comes out 1.1 points above the base, and we get an entire new class at 74.5 AP on top of that. Nothing else about the model changes. Parameters, FLOPs, and inference cost are the same as stock YOLOX-S, and Apache-2.0 carries over from the base, so the weights can be deployed commercially with no per-deployment license and no obligation to open-source derivative work.

Siblings for scale, same recipe and same eval protocol.

Family member	mAP (81 cls)	COCO kept	Cone AP	Inference
yolox_pylon_s	42.0	41.6	74.5	1.7 ms
yolox_pylon_m	47.2	46.8	77.5	2.6 ms
yolox_pylon_l	48.9	48.5	78.6	3.7 ms
yolox_pylon_xl	50.4	50.0	78.8	6.0 ms

We measure inference as forward plus NMS on an A100. Those times are not comparable to the V100 figures published in the official YOLOX table.

Usage

The checkpoint loads with the official YOLOX codebase. The only change from stock YOLOX-S is num_classes = 81, with traffic_cone as class index 80.

import torch
from yolox.exp import get_exp
from yolox.utils import postprocess

# stock yolox-s exp, patched to 81 classes
exp = get_exp(exp_name="yolox-s")
exp.num_classes = 81

model = exp.get_model()
ckpt = torch.load("yolox_pylon_s.pth", map_location="cpu")
model.load_state_dict(ckpt["model"])
model.eval().cuda()

# img is a float32 tensor [1, 3, 640, 640], preprocessed YOLOX-style
with torch.no_grad():
    outputs = model(img)
outputs = postprocess(outputs, num_classes=81, conf_thre=0.25, nms_thre=0.45)

COCO_CLASSES = [...]                        # standard 80-class list
CLASSES = COCO_CLASSES + ["traffic_cone"]   # index 80

Or with the repo's demo tool.

git clone https://github.com/Megvii-BaseDetection/YOLOX && cd YOLOX
python tools/demo.py image \
    -f exps/default/yolox_s.py \
    -c yolox_pylon_s.pth \
    --path your_image.jpg --conf 0.25 --nms 0.45 --tsize 640 --device gpu
# patch exps/default/yolox_s.py with self.num_classes = 81 first

Training

• Base. YOLOX-S (9.0M params), initialized from COCO-pretrained weights
• Data. 147k images across 81 classes, COCO train2017 plus roughly 30k traffic-cone images, trained jointly so the original 80 classes stay in the mix during adaptation
• Eval. COCO val2017 plus a held-out cone split, single 81-class evaluation pass
• Input. 640×640

Intended use and limitations

We built this for roadside and infrastructure perception where traffic cones matter, such as work zones, lane closures, and autonomous driving research, and as a template for class-extension on YOLOX. The S size is the fastest in the family and targets embedded boards and edge cameras.

The model detects boxes for 81 classes. It does not segment, track, or estimate distance. Accuracy on small objects is 24.5 AP, which follows the usual small-model pattern. Where small or distant cones dominate the footage, yolox_pylon_m or yolox_pylon_l is the better choice. As with any detector, we recommend validating on the target cameras before production use.

About Empirisch Tech

We are Empirisch Tech GmbH, a Vienna-based AI company, and we publish YOLOX-Pylon under our Chaperone AI brand. We run one recipe across three domains. We adapt a proven foundation model to a specific domain, keep what the base already knows, and ship the checkpoint together with the data it was trained on.

• Language. Thinking-LQ-1.0 (84% MedQA, within 4 points of GPT-4o at ~20GB) and Coder-LQ-1.0
• Physics. Chaperone-Flow-1.0 (Poseidon-B extended to new CFD regimes, 1.8% wake error) and Palace-LoRA (electromagnetics solver configs)
• Vision. The YOLOX-Pylon family and a road-scene anomaly segmentation pipeline

These models power our production platforms, including NumericalAI (GPU physics simulation) and Simvera (industrial perception trained in simulation, deployed on real cameras). We self-host everything in our own Vienna datacenter and use no third- party model APIs. We are a member of the NVIDIA Inception and Microsoft for Startups programs, and our open checkpoints have passed 30,000 downloads on Hugging Face.

Custom builds. The cone class took one adaptation run. For other classes, cameras, or datasets, reach out via chaperoneai.com/contact.

License

Apache-2.0, matching the YOLOX base.

Citation

@article{yolox2021,
  title={YOLOX: Exceeding YOLO Series in 2021},
  author={Ge, Zheng and Liu, Songtao and Wang, Feng and Li, Zeming and Sun, Jian},
  journal={arXiv preprint arXiv:2107.08430},
  year={2021}
}