Detection benchmarks · YOLOX · 81 classes

A new class, no new cost.

We fine-tuned all four YOLOX tiers on 147k images spanning 81 classes — the 80 COCO classes plus a traffic-cone class annotated entirely by a proprietary model. The result: the extra class lands at the top of the accuracy table, while the original 80 COCO classes stay within about a point of the stock baseline either way, at 1.7–6.0 ms per frame.

Download the weights on Hugging Face ↗

Apache-2.0 · s · m · l · xl

Training setup The lineup vs. baseline Detections Per-class

01 / Training setup

One dataset, four tiers, zero hand labels for the new class.

Every tier was trained on the same corpus: COCO extended with 30k images containing traffic cones. Those cones exist in COCO imagery but carry no annotations — we separated them out and labelled them with a proprietary cone-detection model, turning an unlabelled object into a first-class detection target without a single annotation hour.

All four checkpoints are published under Apache-2.0 on Hugging Face as the yolox-pylon family — s · m · l · xl. Each card carries the full per-class AP table for all 81 classes.

Training images

147k

Extended COCO corpus used identically across all four model tiers.

Classes

The 80 standard COCO classes plus traffic_cone.

Cone images added

30k

Unlabelled cones extracted from COCO and auto-annotated.

Manual annotations

Cone labels produced by a proprietary detection model, not by hand.

02 / The lineup

Pick your point on the curve.

Four checkpoints, one recipe. Two mAP columns, because they answer different questions: what the model scores across all 81 classes, and what it still scores on the 80 original COCO classes alone — the second is the one that compares to a stock COCO model.

Checkpoint	mAP · 80 COCO	mAP · all 81	Traffic cone AP	Inference	Weights
yolox-pylon-s	41.6%	42.0%	74.5%	1.72 ms	download ↗
yolox-pylon-m	46.8%	47.2%	77.5%	2.56 ms	download ↗
yolox-pylon-l	48.5%	48.9%	78.6%	3.71 ms	download ↗
yolox-pylon-xl	50.0%	50.4%	78.8%	6.01 ms	download ↗

Inference is the full pipeline — forward pass plus NMS — at batch 1, 640×640, on an NVIDIA A100. Going from s to xl buys 8.4 points of mAP for 4.3 ms; large objects gain the most (+11.9 AP), the cone class the least (+4.3), since it is already near saturation on the smallest tier. AP50, AP75, recall and the full per-class breakdown for all 81 classes are on each model card.

03 / Against the stock baseline

What the extra class costs.

Each checkpoint against its own stock YOLOX baseline — identical backbone, identical parameters, identical FLOPs, only the data changed. The question that matters is retention: after adding an 81st class, what does the model still score on the original 80?

Tier	Stock YOLOX baseline	Ours · retained 80 COCO	Δ	Traffic cone AP
Small	40.5%	41.6%	↑ 1.1	74.5% · new class
Medium	46.9%	46.8%	− 0.1	77.5% · new class
Large	49.7%	48.5%	↓ 1.2	78.6% · new class
X-Large	51.2%	50.0%	↓ 1.2	78.8% · new class

Small comes out ahead of its baseline, Medium holds parity, Large and X-Large give up about a point — and all four gain a cone detector that scores at or near the top of the class table. Baseline figures are the published stock-YOLOX val2017 numbers; ours use the same protocol and the same 80 classes, so this comparison is like for like. Latency is not — the published baselines were timed on a V100, ours on an A100.

04 / Detections

What it looks like.

The top row is the hard case, real traffic scenes where cones sit among cars, buses, pedestrians and signal heads and every class has to hold at once. Below it, every tier on two kinds of input: a rendered street scene with cones, and an ordinary photograph where the original COCO classes have to keep working. Middle row synthetic, bottom row real.

A busy city intersection: a double-decker bus, six cars, two stop signs, two traffic lights and a traffic cone all detected, the cone at 0.86 confidence — complexcone 0.86

A road closure on a hill street: cones and a barrier beside a queue of parked cars, three traffic lights and a group of pedestrians with handbags, all detected together — complexperson 0.92

A highway overpass shot from across the road: two traffic cones on the far shoulder detected at 0.94 and 0.89, plus three traffic lights against a city skyline — complexcone 0.94

An airport apron under shallow depth of field: airliner at 0.94, two ground service trucks and a small traffic cone off the wingtip detected at 0.86 — complexairplane 0.94

yolox-pylon-s on a synthetic urban street: two traffic cones detected at 0.96 and 0.95 confidence, a parked car at 0.84 — scone 0.96

yolox-pylon-m on a rain-wet synthetic city street: two traffic cones, multiple traffic lights, a bench and a car detected — mcone 0.96

yolox-pylon-l on a synthetic urban intersection with traffic cones detected in the foreground — lcone

yolox-pylon-xl on a wet synthetic boulevard: two traffic cones detected at 0.98 and 0.97 confidence and a bus at 0.96 — xlcone 0.98

yolox-pylon-s on a real office scene: two people, laptop, keyboard and mouse detected — sperson 0.94

yolox-pylon-m on a real baseball scene: batter, catcher, umpire and spectators detected along with a baseball bat and glove — mbat 0.83

yolox-pylon-l on a real tennis court: player, tennis racket, sports ball and a row of chairs detected — lracket 0.91

yolox-pylon-xl on a real office desk: monitor, two keyboards, mouse, laptop, bottle and a shelf of books detected — xllaptop 0.95

05 / Per-class performance

The added class is the best class.

Shown for yolox-pylon-xl. The ranking barely moves between tiers — cones, bears and stop signs lead on all four; hair driers, books and handbags trail on all four. Small, occluded or visually inconsistent objects are the hard ones, exactly as in stock COCO.

Strongest classesAP

Traffic cone78.8

Bear76.4

Cat76.0

Bus75.9

Fire hydrant75.8

Hardest classesAP

Hair drier22.7

Broccoli22.7

Handbag21.2

Apple20.2

Book17.9

The full 81-class AP table for every tier is published on the model cards — s · m · l · xl — along with AP by object size, recall, and the training logs.

Want a class we don't have yet?

The traffic-cone class went from unlabelled pixels to the highest-AP class in the set without a single hand-drawn box. The same pipeline works for your objects.

Apply for Demo →