Yolo Inference

---
name: yolo-inference
description: >
  Use when testing, running, or deploying Ultralytics YOLO inference in Platform or code on images, video, webcams, or streams — Platform Predict and dedicated endpoints, model.predict()/yolo predict, Results API boxes/masks/keypoints, persistent tracking, annotated video, and Solutions such as counting, heatmaps, speed, queues, and parking. For optimized runtime exports, see yolo-export.
---

# Inference, Results API & tracking

## Fastest route: test and deploy in Platform

Every Platform model has a **Predict** tab: upload an image, choose an example, or capture a webcam frame; inference runs automatically and shows the task overlay, summary, raw JSON, and timing. Adjust `conf`, `iou`, and `imgsz` with the same meanings used below.

For production, open **Deploy**, choose a nearby region, and wait for the dedicated endpoint to become **Ready**. Its deployment page provides health, metrics, logs, a Predict panel, and ready-to-use Python/JavaScript/cURL examples. Call `/predict` with the deployment's bound API key as a bearer token:

```python
import requests

with open("image.jpg", "rb") as image_file:
    response = requests.post(
        "https://YOUR_DEPLOYMENT_URL.run.app/predict",
        headers={"Authorization": "Bearer YOUR_API_KEY"},
        files={"file": image_file},
        data={"conf": 0.25, "iou": 0.7, "imgsz": 640},
    )
response.raise_for_status()
print(response.json())
```

Default-size endpoints scale to zero, so expect a cold start after idle periods; custom sizes keep a warm instance billed for uptime. See [Platform Inference](https://docs.ultralytics.com/platform/deploy/inference) and [Dedicated Endpoints](https://docs.ultralytics.com/platform/deploy/endpoints).

## Quickstart

```python
from ultralytics import YOLO

model = YOLO("yolo26n.pt")  # or your runs/detect/train/weights/best.pt
results = model("image.jpg")  # list[Results], one per image
results[0].show()
```

```bash
yolo predict model=yolo26n.pt source="image.jpg" save=True
```

**The one rule for video/streams: `stream=True`.** The default builds a list of ALL results in RAM — OOM on long videos. `stream=True` returns a generator:

```python
for r in model("video.mp4", stream=True):
    ...
```

## Sources

Accepted directly: image/video path, directory, glob, URL, webcam index (`0`), RTSP/RTMP/HTTP streams, YouTube URL (needs `pytubefix`), PIL image, numpy array (assumed BGR), torch tensor, or a list of these. `vid_stride=N` processes every Nth frame.

## Arguments that matter

| Arg            | Default | Notes                                                                                                                        |
| -------------- | ------- | ---------------------------------------------------------------------------------------------------------------------------- |
| `conf`         | 0.25    | lower → more recall + more false positives                                                                                   |
| `iou`          | 0.7     | NMS threshold; ignored by the NMS-free head (`nms=False`)                                                                    |
| `nms`          | None    | default: one-to-many head + NMS; `False`: faster NMS-free YOLO26 head; `True`: embed NMS on export. Exports keep that choice |
| `imgsz`        | model   | inherited from the checkpoint; set explicitly for a different shape (static exports always run at their export size)         |
| `classes`      | None    | keep only these ids, e.g. `classes=[0]`                                                                                      |
| `max_det`      | 300     | raise for dense scenes                                                                                                       |
| `quantize`     | None    | `16` selects FP16 compute for PyTorch/TorchScript; artifact/runtime precision otherwise applies; see caveat below            |
| `batch`        | 1       | >1 speeds up folders/videos with `stream=True`                                                                               |
| `retina_masks` | False   | full-resolution masks (slower, crisper)                                                                                      |
| `augment`      | False   | detect-only test-time augmentation (+accuracy, ~3× slower); `nms=False` and exports warn and run normal inference            |
| `verbose`      | True    | False in loops to silence per-frame logs                                                                                     |

`quantize=16` also rounds OpenVINO inputs to FP16. It replaces the deprecated `half` argument.

Saving/drawing: `save`, `save_txt`, `save_conf`, `save_crop`, `show`, `line_width` → `runs/<task>/predict*/`.

## Results API

Each `Results` has the task's payload — `.boxes`, `.masks`, `.keypoints`, `.probs` (classify), `.obb`, `.semantic_mask`, `.depth` — plus `.names` (id→name), `.orig_img` (BGR), `.speed`, and methods `.plot()`, `.show()`, `.save()`, `.save_txt()`, `.save_crop()`, and `.summary()`. The only exporters are `to_df()`, `to_csv()`, and `to_json()`.

```python
r = results[0]
r.boxes.xyxy / .xywh / .xyxyn / .xywhn   # (N,4) boxes, pixel / normalized
r.boxes.conf, r.boxes.cls, r.boxes.id    # confidence, class ids, track ids (track mode)
r.masks.xy                                # list of (K,2) pixel polygons per instance
r.keypoints.xy, r.keypoints.conf          # (N,K,2), (N,K)
r.probs.top1, r.probs.top1conf            # classify; r.names[r.probs.top1] → label
r.obb.xyxyxyxy, r.obb.xywhr               # oriented boxes

from collections import Counter
counts = Counter(r.names[int(c)] for c in r.boxes.cls)   # count per class
```

`model.embed("image.jpg")` returns feature vectors (similarity search, clustering).

## Tracking (persistent IDs across frames)

```python
for r in model.track("video.mp4", stream=True):
    ids = r.boxes.id  # tensor of track ids, or None

# Frame-by-frame loop with your own capture: persist=True is REQUIRED
r = model.track(frame, persist=True)[0]  # else the tracker resets every frame
```

- Six trackers, selected with `tracker=`: **`tracktrack.yaml` (default)**, `botsort.yaml`, `bytetrack.yaml`, `ocsort.yaml`, `deepocsort.yaml`, `fasttrack.yaml`. ByteTrack is lightest; BoT-SORT/TrackTrack/DeepOCSORT support ReID (`with_reid: True`) and camera-motion compensation (`gmc_method`).
- Custom behavior: copy the YAML from `ultralytics/cfg/trackers/`, tweak (`track_buffer` = frames a lost track survives, `track_high_thresh`), pass your path.
- Tracking accepts all predict args but defaults to `conf=0.1` so trackers see low-confidence boxes. Always guard `r.boxes.id is not None`.

## Annotated-video writing pattern

```python
import cv2

cap = cv2.VideoCapture("in.mp4")
w, h, fps = (int(cap.get(p)) for p in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS))
out = cv2.VideoWriter("out.mp4", cv2.VideoWriter_fourcc(*"mp4v"), fps, (w, h))
for r in model.track("in.mp4", stream=True):
    out.write(r.plot())  # .plot() returns annotated BGR frame
cap.release()
out.release()
```

## Performance checklist

1. For PyTorch/TorchScript on a supported GPU, benchmark `quantize=16` against FP32.
2. Export to the target-native backend (TensorRT/OpenVINO/CoreML) and benchmark it (see yolo-export; exports load straight back into `YOLO()`).
3. Use a smaller model or `imgsz`, or the NMS-free YOLO26 head (`nms=False`).
4. Set `batch>1` for offline folders; use `vid_stride` when every frame isn't needed.
5. Set `verbose=False`; skip `.plot()` when only coordinates are needed.
6. Use one `YOLO()` instance per thread — never share across threads.

## Troubleshooting

| Symptom                            | Cause / fix                                                                                                                                   |
| ---------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------- |
| No detections on visible objects   | `conf` too high; wrong weights; `imgsz` far from training size                                                                                |
| Boxes offset                       | you pre-resized manually — pass the raw image, preprocessing is internal                                                                      |
| Red/blue swapped in PIL/matplotlib | Results arrays (`orig_img`, `.plot()`) are BGR; convert with `cv2.cvtColor(..., cv2.COLOR_BGR2RGB)`                                           |
| RAM climbs on video                | missing `stream=True`                                                                                                                         |
| `boxes.id is None` crash           | guard for None; `persist=True` in manual loops                                                                                                |
| Duplicate boxes                    | lower `iou`; drop `nms=False` (its NMS-free head ignores `iou`; re-export such artifacts); add `agnostic_nms=True` for cross-class duplicates |
| Caps at 300 objects                | raise `max_det`                                                                                                                               |
| Slow first inference               | warmup — benchmark from the second call                                                                                                       |

## Related pages

- `solutions.md` (this folder) — read BEFORE hand-rolling counting, heatmaps, speed estimation, zone logic, queues, parking, workout counting, or privacy blurring: a prebuilt Solution almost certainly exists.

If the installed version rejects an argument, trust the error text and `yolo cfg` over this file.