Choose and benchmark the detector and tracker

Two components decide how well CV-Scope sees:

The detector finds objects in each frame. Larger detectors find more, especially small or distant objects, but need more computing power.
The tracker links the detections of consecutive frames into one track per object. It decides whether a person keeps one track number from entering to leaving.

Step 1. See what this computer can run

Open Hardware. It lists the processor, memory, graphics card and the acceleration CV-Scope can use, such as CUDA on NVIDIA cards or Metal on Apple computers.
Recommended detectors for this system lists four tiers: Fast, Balanced, High accuracy and CPU mode. The one marked recommended is used when an experiment is left on automatic.
Click Probe again after installing a new graphics driver or card.

Step 2. Install detectors

Open Models. The Detector catalog lists each detector with its size, the memory it needs, dots for relative accuracy and speed, and its licence.
Click Install on the detectors you want to compare. The Status column shows the download and then installed.
Click a detector’s name for details, including the classes it knows.

Step 3. Benchmark on your own footage

A benchmark runs about 120 frames of the most recently added video through the detector and the tracker, and measures the time each part takes.

On Models, click Benchmark next to an installed detector.
Keep Device on Auto. Leave Inference size (px) empty to use the detector’s default.
Choose the Tracker you plan to use, and click Run benchmark.
The result appears under Benchmarks on this machine. Compare Pipeline FPS, the frames per second of the whole chain.

As an example, YOLO11n with ByteTrack ran at about 130 frames per second on a desktop with a mid-range NVIDIA card. The same detector on a laptop processor runs far slower.

Rule of thumb. Pick the largest detector whose Pipeline FPS is at least twice the processing rate you need for each camera running at the same time. For three cameras at 5 frames per second each, that is 30 or more.

The same benchmark runs from the command line:

.venv\Scripts\python.exe -m pathscope.cli benchmark yolo11n --video samples\videos\people-detection.mp4

Step 4. Set the detector of an experiment

On the experiment page, in Detection and tracking:

Setting	What it does
Detector	Recommended for this system (AUTO) follows the hardware recommendation. Or pick an installed detector.
Preset	With AUTO, picks the tier: FAST, BALANCED or ACCURATE. CUSTOM opens every setting.
Tracker	ByteTrack or BoT-SORT.
Playback	Paces a video file at its real speed, for a live feel. Off processes as fast as possible.

When a preset asks for a detector that is not installed, the run uses another installed detector from the table on Hardware. Its results page says so under Diagnostics, in the Configuration snapshot, on the resolution line. If none of those detectors is installed, the run does not start. AUTO and the presets follow the hardware recommendation only while Default detector on Settings is empty.

Tick Advanced settings for the details:

Setting	When to change it
Device	Force the processor or the graphics card.
Inference resolution (px)	Larger finds small, distant objects, at the cost of speed. When it is empty, the camera’s own Inference resolution (px) applies, then the detector’s default.
Confidence threshold	Lower finds more but also more false objects. The default is 0.25.
Box overlap threshold (IoU)	Controls when two overlapping boxes count as one object.
Processing FPS	Frames analysed per second. Fewer saves processing.
Frame skipping	Skips frames after each processed frame.

Step 5. Choose the tracker

ByteTrack is fast and works well for a fixed camera. Keep it unless you see one of the problems below.

BoT-SORT adds two things. Set Tracker to BoT-SORT and tick Advanced settings to see them under BoT-SORT settings:

Compensate for camera movement. For a camera that shakes in the wind or gets bumped, it measures how the whole picture moved and corrects the tracks. Optical flow (recommended) suits most cases. Choose Off (the camera never moves) for a fixed indoor camera.
Use appearance to keep identities through occlusions. When people cross behind each other, appearance helps to keep their track numbers apart:
- Colour histogram (fast, no model) compares clothing colours.
- Deep features, ResNet-18 (GPU recommended) is more reliable. Install ResNet-18 appearance features on Models first.
- A custom re-identification model in ONNX format can be placed in the reid folder of the models directory, data\models\reid by default. It then appears in the list.

Appearance summaries are held in memory only, while an object is tracked. They are never stored and never compared across runs.

The tracker settings that matter most:

Frames to keep searching for a lost object. Counted in processed frames. Raise it when people disappear behind a pillar for a while. Too high and two different people can get the same track.
Frames before a new object is reported. Raise it when short false detections create tracks that count.
Detection confidence to start or continue a track. The tracker starts a track only from a detection at least this confident, 0.5 by default. When objects are missed, lower it to about 0.3 together with the detector’s Confidence threshold. BoT-SORT has a separate Detection confidence to start a new track.

Step 6. Check the choice

Settings are only better when the results are better. Compare two experiments that differ only in the detector or tracker, as in guide 14. Then judge a sample of events for each, as in guide 17.

Guide 20 sends events to other tools as they happen, and reads results through the API.