YOLOX-Pylon-M

A YOLOX-M 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 this checkpoint holds the official YOLOX-M baseline to within 0.1 mAP, 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 accuracy and speed balance point of 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)	47.2
mAP 50:95, original 80 COCO classes only	46.8
AP50 / AP75	64.9 / 51.3
AP small / medium / large	29.0 / 51.7 / 61.4
AR@100	60.2
traffic_cone AP / AR	77.5 / 80.4
Inference (forward + NMS, batch 1, A100)	2.56 ms

Two things stand out.

• Retention is essentially free. The official YOLOX-M val2017 baseline is 46.9 mAP on COCO. After adding the cone class, this checkpoint keeps 46.8 on the same 80 classes, a 0.1-point cost for an entire new class.
• The added class is the best class. At 77.5 AP, traffic_cone outscores every one of the 80 original classes on this checkpoint. The next best are bear (76.6), stop sign (75.3), and fire hydrant (74.2).

Full per-class AP (81 classes)

class	AP	class	AP	class	AP
person	59.995	bicycle	35.488	car	56.375
motorcycle	48.854	airplane	70.677	bus	73.185
train	70.605	truck	51.140	boat	31.545
traffic light	41.793	fire hydrant	74.168	stop sign	75.333
parking meter	49.909	bench	33.664	bird	40.498
cat	70.232	dog	64.979	horse	66.330
sheep	56.374	cow	60.443	elephant	70.077
bear	76.568	zebra	73.018	giraffe	72.611
backpack	19.293	umbrella	46.162	handbag	18.271
tie	38.489	suitcase	45.225	frisbee	69.493
skis	30.339	snowboard	42.471	sports ball	48.027
kite	47.608	baseball bat	34.778	baseball glove	40.530
skateboard	59.871	surfboard	42.286	tennis racket	55.113
bottle	42.121	wine glass	38.985	cup	45.900
fork	43.635	knife	24.711	spoon	23.156
bowl	45.258	banana	28.749	apple	22.998
sandwich	35.454	orange	28.775	broccoli	24.286
carrot	26.312	hot dog	43.187	pizza	56.291
donut	51.690	cake	41.412	chair	37.081
couch	50.049	potted plant	31.895	bed	47.375
dining table	33.759	toilet	65.277	tv	61.131
laptop	64.677	mouse	61.303	remote	35.245
keyboard	53.938	cell phone	38.263	microwave	66.421
oven	40.868	toaster	41.509	sink	39.656
refrigerator	61.356	book	15.539	clock	50.358
vase	40.381	scissors	37.710	teddy bear	49.235
hair drier	9.187	toothbrush	31.041	traffic_cone	77.503

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_m (this model)	46.8 kept + traffic_cone 77.5	25.3M	73.8G	cone added, COCO kept	Apache-2.0
YOLOX-M (base)	46.9	25.3M	73.8G	COCO only	Apache-2.0

Reading that table, the adaptation costs us 0.1 mAP on the original 80 classes, which sits inside the noise, and buys an entire new class at 77.5 AP. Nothing else about the model changes. Parameters, FLOPs, and inference cost are the same as stock YOLOX-M, 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-M 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-m exp, patched to 81 classes
exp = get_exp(exp_name="yolox-m")
exp.num_classes = 81

model = exp.get_model()
ckpt = torch.load("yolox_pylon_m.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_m.py \
    -c yolox_pylon_m.pth \
    --path your_image.jpg --conf 0.25 --nms 0.45 --tsize 640 --device gpu
# patch exps/default/yolox_m.py with self.num_classes = 81 first

Training

• Base. YOLOX-M (25.3M 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 M size is the accuracy and speed balance point for most fixed-camera deployments.

The model detects boxes for 81 classes. It does not segment, track, or estimate distance. Where small or distant objects dominate the footage, at 29.0 AP on the small bucket, yolox_pylon_l is worth considering. For embedded and edge boards, yolox_pylon_s runs in 1.7 ms. 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}
}