Backbone, neck, head
Every modern detector has this three-part shape, and it is worth being able to name the parts.
The backbone is a classification network with the classifier removed — in v8, a CSPDarknet variant whose repeating unit is the C2f block: a split, several bottleneck convolutions, and a concatenation of all the intermediate outputs. It produces feature maps at strides 8, 16 and 32.
The neck mixes those levels. Semantic information is strongest at stride 32 and spatial precision is strongest at stride 8, so a top-down path carries semantics down and a bottom-up path carries precision back up — the PAN arrangement, itself a descendant of FPN. Without it, the level that can see small objects has no idea what they are.
The head turns each level into predictions. v8's head is *decoupled*: one small convolutional branch for classification and a separate one for box regression, rather than a single branch predicting both. Sharing them makes the two tasks fight over the same features, and separating them is worth about a point of mAP for a small amount of compute.
Anchor-free, and what that replaced
Versions 2 through 7 were anchor-based: each cell carried several prior boxes of fixed size and aspect ratio, and the network predicted an offset from the nearest one. That meant a set of anchor dimensions had to be chosen per dataset — usually by k-means over the training boxes — and it meant three or more predictions per cell.
v8 predicts, from each cell, the distances from that cell's centre to the four edges of the box. No priors, no offsets, one prediction per cell. The hyperparameter is gone.
The regression is not a plain regression, though. Each of the four distances is predicted as a distribution over 16 bins, so the box branch emits 4 × 16 = 64 channels per cell, and the actual distance is the expectation of that distribution. This is Distribution Focal Loss, and the reason for it is that box edges are genuinely ambiguous — where exactly does a blurred or occluded boundary lie? — and a distribution can express "probably 12, possibly 15" where a single number cannot. The network's own uncertainty becomes trainable.
There is also no objectness score in v8. Older versions predicted "is there anything here" separately from "what is it"; v8's confidence is just the class score. One fewer output, one fewer loss term, one fewer thing to calibrate.
Assignment during training is handled by TaskAlignedAssigner: rather than a fixed rule about which cell owns which object, it scores each candidate by a combination of its classification confidence and its IoU with the ground truth, and assigns the top few. Cells that are already good at both get the label, which is a mild form of the network choosing its own supervision.
The two thresholds
Everything above is fixed at training time. The two controls in the explorer are not — they are applied after the forward pass, and changing them changes your reported metrics without retraining anything.
Confidence threshold. Every prediction below it is discarded. Push it up in the explorer and boxes vanish; push it past 0.55 and the note changes to warn you, because the objects you lose first are exactly the hard ones — small, occluded, unusual pose — which are the ones with low scores. A high threshold makes a demo look clean and makes recall quietly terrible. The default of 0.25 for visualisation is a display choice; for computing mAP you use something like 0.001, because mAP integrates over the whole precision-recall curve and truncating it early just throws away area.
NMS IoU threshold. Several neighbouring cells will each decide the same object is theirs, so the head emits a cluster of near-identical boxes. Non- maximum suppression sorts by confidence, keeps the top box, and deletes anything overlapping it by more than the threshold, per class. Turn NMS off in the explorer to see the raw cluster.
The threshold is a genuine trade and the "crowd" scene is there to show it. Set it low, around 0.3, and two people standing close — who really do overlap by more than 0.3 — are treated as duplicates and one is deleted. Set it high, around 0.9, and duplicates survive. There is no value that is right for both a sparse scene and a dense one, which is why crowded-scene detection has its own literature: Soft-NMS, which decays scores instead of deleting, and set-prediction models like DETR, which have no NMS at all because their loss forbids duplicate predictions in the first place.
from ultralytics import YOLO
model = YOLO("yolov8n.pt")
results = model.predict(
"street.jpg",
conf=0.25, # the confidence cut
iou=0.7, # the NMS IoU threshold
imgsz=640, # must be a multiple of 32: strides are 8, 16, 32
max_det=300,
)
for box in results[0].boxes:
print(model.names[int(box.cls)], float(box.conf), box.xyxy[0].tolist())
imgsz has to be a multiple of 32 or the three levels do not divide evenly. Raising it improves small-object recall roughly in proportion to the extra pixels and costs latency quadratically — usually the single most effective knob, and the one people forget exists while they tune conf.
Why the model sizes are letters
v8 ships as n, s, m, l and x — nano through extra-large — and they are the same architecture with two multipliers applied: one to the channel counts and one to the number of repeats in each C2f block. Nothing structural differs between them.
That matters because it makes the accuracy/latency trade a dial rather than a choice of architecture, and because the scaling is predictable: roughly, depth and width both scale, so parameters grow faster than latency does. The nano model is around 3 M parameters and the extra-large around 68 M, for perhaps five points of COCO mAP between them.
The practical sequence for picking one is the reverse of what people usually do. Start from the latency budget on the actual target hardware, pick the largest model that fits it, and only then look at accuracy. Choosing the model first and optimising afterwards usually ends in quantisation and pruning work that a smaller model would have made unnecessary.
Reading the mismatch counter
The explorer's fourth statistic is missed / duplicate / false, recomputed as you move the sliders. It is worth playing with deliberately, because the three failure modes have three different causes:
- Missed objects mean the confidence threshold is too high, or the object is smaller than P3 can resolve.
- Duplicates mean the NMS threshold is too high, or NMS is off.
- False positives mean the confidence threshold is too low.
Every real tuning session is spent trading these against each other, and mAP is the single number that summarises the whole curve so you do not have to pick a point on it until deployment.