YOLO27 Crosses 60 mAP: What the Next Detector Generation Means for CCTV Games

Ultralytics used its YOLO Vision 2026 stage on 13 September to unveil YOLO27, and the headline number is one CCTV Games watchers will recognise. YOLO27l is the first Ultralytics model to pass 60 mAP on COCO. Earlier this year that mark belonged to RF-DETR alone. Now the two biggest real-time detector families sit on the same side of the line, and the next fight moves to the parts of the frame that decide a count.

What Ultralytics announced

The Ultralytics event recap describes YOLO27 as 28 models across four sizes (n, s, m and l) and seven tasks, from detection and segmentation to depth estimation and pose. Ultralytics also previewed domain-specific pretrained models for retail, warehousing, security and aerial vision.

The lineup is simpler than before. In its launch post, Ultralytics said n and s are for edge devices, m for demanding real-time detection, and l for "high-precision and safety-critical systems". There is no x model, because the l model "already delivers flagship accuracy". The company claims higher accuracy across the range "on the same hardware and at the same speed".

Alongside it came YOLOE-27, an open-vocabulary model that detects objects from text, visual prompts or no prompt at all. Both are listed as "coming soon" on the Ultralytics roadmap, with release due later this year.

Why 60 mAP matters less than it sounds

COCO mAP is an average across 80 object classes and many box sizes. A CCTV Game cares about one class per round: vehicles in Rush Hour, ducks in Duck River, skiers in Snow Run. What decides the integer is narrower:

• Small, distant objects. A car at the back of a junction is a few dozen pixels wide.
• Occlusion. A bus hiding a motorbike for half a second is the classic miscount.
• Night and weather. Headlight glare, rain and snow glare change what the camera sees.
• Tracking. The detector finds the car. The tracker decides it is the same car on frame 12 and frame 47. We covered that layer in our tracking engine piece.

Ultralytics has not published YOLO27 numbers for any of those four. Until independent benchmarks land, the 60 mAP claim is a signal about direction, not proof of better counts.

Tracking research keeps moving too

The detector is only half the story. ECCV 2026 in September included new tracking work such as SelfMOTR, which revisits end-to-end tracking with self-generated detection priors. On 7 October, MOTIP2 appeared on arXiv with spatial priors for end-to-end multi-object tracking. End-to-end trackers try to fold detection and identity into one model, which is exactly where occlusion errors live.

What it means for players

155.io has not publicly confirmed a move to YOLO27, and nothing here suggests a switch is coming. So treat this as context, not a product update. Three things are worth knowing:

• Better models narrow disputes, they do not change odds. 155.io publishes the payout and RTP for every selection, and those are fixed per bet in its payout reference.
• Edge speed matters for latency. Faster small models mean counts can be computed closer to the camera, which helps the gap between what you see and what settles.
• Open-vocabulary detection widens the menu. A model that can count "anything you can describe" makes new CCTV formats cheaper to build. That is a long-run point, not a launch.

The bottom line

YOLO27 puts the YOLO family back level with RF-DETR on the benchmark that gets the headlines. The benchmark that matters for CCTV Games is whether the boxes on your screen match the count that settles. Watch for independent tests on crowded, low-light street scenes once the models ship.