Figure 1 — Object size distribution in the VisDrone training set (343,205 objects across 6,471 images). Left: histogram of geometric mean object size. 51.8% of objects fall below the 32px threshold, the point below which detection becomes unreliable at standard 640px resolution. Right: cumulative distribution confirming the majority of VisDrone objects are in the small and very small range.

The most direct fix for the downsampling problem is to use a higher input resolution. The Ultralytics default is 640×640, which is the right starting point for most detection tasks. But understanding what resolution does to your specific objects is more important than changing it blindly.

What Resolution Actually Does

At input resolution R, an object of width w pixels occupies w/R fraction of the image. At the deepest feature map (stride 32), that object is represented by w/32 feature map cells. For an object to be reliably detected, it needs to occupy at least 1 to 2 feature map cells — meaning its pixel width should be at least 32 pixels at the input resolution.

To demonstrate the impact of resolution on small object detection, we trained identical YOLOv12n models at 320px and 640px on VisDrone and evaluated on the same validation set. The results are unambiguous:

Figure 2 — YOLOv12n predictions on the same VisDrone validation image at 320px (left, red boxes) and 640px (right, green boxes). At 320px the model detects fewer objects and misses many small vehicles entirely. At 640px the same model finds significantly more objects with tighter localisation. Same confidence threshold (0.25) applied to both.

The mAP50 difference between 320px and 640px is 0.148, an 84% relative improvement from doubling the input resolution alone, with no other changes. This is the single largest gain of any strategy tested.

from ultralytics import YOLO

model = YOLO('yolo12n.pt')

# Low resolution — inadequate for small objects
model.train(
    data='visdrone.yaml',
    imgsz=320,
    batch=16,
    epochs=150,
    patience=20,
    name='yolo12n_320'
)

# Default resolution — significantly better
model.train(
    data='visdrone.yaml',
    imgsz=640,
    batch=8,
    epochs=150,
    patience=20,
    name='yolo12n_640'
)

Resolution addresses the downsampling problem but does not fully solve the feature quality problem. A larger model with more parameters produces richer feature representations at every scale — which matters most for small objects where the available signal is already sparse.

What Model Scale Changes

YOLOv12 comes in five scales, nano (n), small (s), medium (m), large (l), and extra-large (x). Moving from nano to small increases the feature pyramid channel width, which strengthens feature representations at the small-object detection head. The improvement for small objects is proportionally larger than for large objects because the wider channels provide more capacity to represent the subtle, low-signal features of small targets.

We trained YOLOv12s at 640px, the same resolution as the nano baseline, and measured the improvement:

from ultralytics import YOLO

# Nano baseline
model_n = YOLO('yolo12n.pt')
model_n.train(
    data='visdrone.yaml',
    imgsz=640,
    batch=8,
    epochs=150,
    patience=20,
    name='yolo12n_640'
)

# Small model — stronger features
model_s = YOLO('yolo12s.pt')
model_s.train(
    data='visdrone.yaml',
    imgsz=640,
    batch=8,
    epochs=150,
    patience=20,
    name='yolo12s_640'
)

The first two strategies require retraining. This one does not. NMS threshold adjustment applies to an already-trained model at inference and evaluation time, it takes minutes rather than hours and can recover meaningful performance with no compute cost.

How NMS Affects Small Object Detection

During inference, after the model produces raw predictions, Non-Maximum Suppression removes duplicate detections. The NMS IoU threshold controls how much overlap two predictions must have before one is suppressed. The Ultralytics default is iou=0.7 for training-time evaluation.

For dense small object scenes, two adjacent objects, pedestrians walking closely, vehicles in tight parking rows, may have bounding boxes that overlap significantly. At iou=0.7, these are kept correctly. But at inference, the suppression behaviour interacts with prediction confidence distributions in ways that can affect recall for small objects specifically.

We evaluated the same YOLOv12s 640px checkpoint across three NMS IoU thresholds using the standard evaluation protocol (conf=0.001) to get comparable mAP numbers:

The sweet spot is iou=0.5 , a +0.005 mAP50 improvement and +0.007 recall improvement over the default, with mAP50-95 unchanged. Going lower to 0.3 shows diminishing returns, the more aggressive suppression starts hurting precision without further recall gains.

The gain is modest but completely free, no retraining, no additional compute, just a single parameter change at evaluation or inference time:

from ultralytics import YOLO

model = YOLO('yolo12s_640/weights/best.pt')

# Default NMS threshold
results_default = model.val(
    data='visdrone.yaml',
    imgsz=640,
    conf=0.001,
    iou=0.7          # default
)

# Adjusted NMS threshold — better for small objects
results_adjusted = model.val(
    data='visdrone.yaml',
    imgsz=640,
    conf=0.001,
    iou=0.5          # sweet spot for VisDrone
)

# Apply at inference too
predictions = model.predict(
    source='visdrone/images/val',
    conf=0.25,
    iou=0.5,         # same adjustment at inference
    save=True
)

Figure 3 — mAP50 and mAP50-95 on VisDrone validation set across three improvement strategies applied progressively to YOLOv12. Resolution produces the largest single gain (+0.148). Model scale provides a meaningful secondary improvement (+0.089). NMS threshold adjustment adds a small but free final gain (+0.005). All results use the standard evaluation protocol (conf=0.001, iou as specified).

Companion Resources Included

• Resolution training configuration
• Model scale comparison configuration
• NMS threshold tuning notes
• Minimum object size reference
• Small object detection checklist
• Deployment trade-off guide

Working on a detection project with small objects and running into challenges this post does not fully address?
Reach out through Occulins Contact for deployment-aware computer vision research and engineering support.