YOLOX-Pylon-L

A YOLOX-L 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 retains 98% of the official YOLOX-L baseline while the added cone class scores higher than 79 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 strongest checkpoint that still runs comfortably in real time.

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)	48.9
mAP 50:95, original 80 COCO classes only	48.5
AP50 / AP75	66.0 / 52.7
AP small / medium / large	31.1 / 53.2 / 63.5
AR@100	60.8
traffic_cone AP / AR	78.6 / 81.4
Inference (forward + NMS, batch 1, A100)	3.71 ms

Two things stand out.

• Retention held at 98%. The official YOLOX-L val2017 baseline is 49.7 mAP on COCO. After adding the cone class, this checkpoint keeps 48.5 on the same 80 classes, so we traded 1.2 points for an entire new class arriving at top-tier accuracy.
• The added class is the best class. At 78.6 AP, traffic_cone is second only to bear (79.7) on this checkpoint, ahead of giraffe (76.4) and stop sign (75.8), and far ahead of workhorse classes like person (61.6) and car (58.5).

Full per-class AP (81 classes)

class	AP	class	AP	class	AP
person	61.618	bicycle	37.547	car	58.494
motorcycle	51.265	airplane	73.783	bus	74.632
train	71.065	truck	53.815	boat	30.934
traffic light	43.080	fire hydrant	75.738	stop sign	75.843
parking meter	51.456	bench	37.430	bird	41.997
cat	74.731	dog	67.190	horse	67.541
sheep	59.677	cow	63.133	elephant	71.051
bear	79.689	zebra	74.188	giraffe	76.406
backpack	20.183	umbrella	47.879	handbag	20.090
tie	43.556	suitcase	48.253	frisbee	69.678
skis	33.713	snowboard	41.683	sports ball	50.590
kite	51.097	baseball bat	39.119	baseball glove	41.565
skateboard	62.194	surfboard	45.487	tennis racket	59.014
bottle	42.903	wine glass	41.440	cup	48.639
fork	49.134	knife	26.992	spoon	26.324
bowl	46.760	banana	28.472	apple	19.796
sandwich	35.887	orange	31.618	broccoli	24.728
carrot	27.921	hot dog	43.810	pizza	59.750
donut	50.739	cake	43.368	chair	38.982
couch	49.349	potted plant	30.830	bed	49.074
dining table	30.917	toilet	64.619	tv	61.952
laptop	68.424	mouse	64.775	remote	43.296
keyboard	56.172	cell phone	44.216	microwave	69.574
oven	43.304	toaster	28.361	sink	43.081
refrigerator	62.442	book	16.598	clock	52.497
vase	41.479	scissors	34.524	teddy bear	52.537
hair drier	10.184	toothbrush	29.280	traffic_cone	78.623

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_l (this model)	48.5 kept + traffic_cone 78.6	54.2M	155.6G	cone added, COCO kept	Apache-2.0
YOLOX-L (base)	49.7	54.2M	155.6G	COCO only	Apache-2.0

Reading that table, the adaptation costs us 1.2 mAP on the original 80 classes and buys an entire new class at 78.6 AP. Nothing else about the model changes. Parameters, FLOPs, and inference cost are the same as stock YOLOX-L, 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-L 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-l exp, patched to 81 classes
exp = get_exp(exp_name="yolox-l")
exp.num_classes = 81

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

Training

• Base. YOLOX-L (54.2M 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 L size is the accuracy pick that still runs comfortably in real time, and it fits server- side inference over multiple camera streams as well as offline analysis and auto-labeling. It also holds the family best on small objects at 31.1 AP, ahead of the larger XL checkpoint.

The model detects boxes for 81 classes. It does not segment, track, or estimate distance. For fixed single-camera deployments, yolox_pylon_m offers most of the accuracy at 2.6 ms. 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}
}