Console Vision Glossary: 58 Capture, Vision and Input Terms

0–9 and A–C: 1080p120 to CUDA

Terms are grouped alphabetically. Product and brand names belong to their owners; IRONHIDE Software is not affiliated with ConsoleTuner, Sony, Microsoft, NVIDIA, Elgato, XIM or any other maker named here.

• 1080p120 — A video mode of 1920 × 1080 pixels at 120 frames per second. PS5 and Xbox Series X|S can output 120 Hz, but not every capture card can capture 1080p120 rather than only pass it through. At 120 fps a vision pipeline has 8.3 ms per frame.
• Bounding box — The rectangle an object detector places around each object, stored as x, y, width and height (or two corner points) in pixels. Every box comes with a class label and a confidence score.
• Capture card — A device that takes a console’s HDMI output and delivers it to a PC as a video source, usually over USB. See our capture card buying guide.
• Classification — The computer-vision task that assigns one label to a whole image, such as “menu” or “gameplay,” with a confidence score.
• CNN (convolutional neural network) — A neural network that slides small learned filters across an image to detect edges, textures and shapes. Real-time detectors such as the YOLO family are built mainly from convolutional layers.
• Confidence score — A number between 0 and 1 that expresses how sure a model is about a prediction. Results below a threshold, often 0.25 for detectors, are dropped.
• CUDA — NVIDIA’s GPU computing platform. ONNX Runtime’s CUDA execution provider runs models on NVIDIA GPUs and needs CUDA and cuDNN versions that match the ONNX Runtime build.

D–G: Dataset to Gtuner IV

• Dataset — A collection of labeled examples, such as frames with boxes drawn around the objects of interest, used to train or evaluate a model.
• DigitalXIM — An IRONICS plugin that sends native keyboard and mouse input from the PC to the console, replacing a hardware XIM adapter. It works with Titan Two, Makcu and Remote Play, alongside the other modules or completely standalone without computer vision. See what DigitalXIM is.
• DirectML — Microsoft’s low-level machine-learning API for DirectX 12 GPUs from NVIDIA, AMD, Intel and Qualcomm. ONNX Runtime uses it through the DirectML execution provider. DirectML is in sustained engineering: still supported, with new feature work moved to Windows ML.
• Execution provider (EP) — An ONNX Runtime back end for one kind of hardware, such as CPU, DirectML, CUDA or TensorRT. You list EPs in priority order, and each part of the model runs on the first EP that supports it. Details in our ONNX Runtime and DirectML guide.
• FP16 / FP32 — 16-bit and 32-bit floating-point numbers. FP32 is the default precision for model weights; converting to FP16 cuts model size by up to half and can speed up GPU inference at a small accuracy cost.
• Frame rate (fps) — How many frames per second a video source produces or a pipeline processes. Console games commonly run at 30, 60 or 120 fps.
• Frame time — The time per frame: 1,000 ms divided by the frame rate. That is 33.3 ms at 30 fps, 16.7 ms at 60 fps and 8.3 ms at 120 fps, and it is the budget a real-time pipeline must stay under.
• GPU (graphics processing unit) — A graphics card or integrated graphics chip. Its parallel design makes it much faster than a CPU for neural-network inference.
• Gtuner IV — ConsoleTuner’s software for programming, updating and configuring the Titan Two, including loading Input Translators into memory slots.

H–I: HDCP to IRONICS

• HDCP (High-bandwidth Digital Content Protection) — Copy protection on HDMI links. Capture cards cannot use HDCP-protected signals and show a black image; on a PS5, turn off Enable HDCP in the System settings before capturing.
• HDMI 2.1 — The HDMI version that adds the bandwidth for modes such as 4K at 120 Hz and gaming features such as VRR. A capture card’s passthrough must support these modes if you want to keep them on your TV.
• HDR (high dynamic range) — A video format with a wider brightness and color range than SDR. Capture cards list HDR passthrough and HDR capture as separate features, often with different resolution limits.
• HID (Human Interface Device) — The USB device class for keyboards, mice, game controllers and similar input devices. HID devices describe their own data reports, so a host can use them with generic drivers.
• Inference — Running a trained model on new input to get predictions, as opposed to training it. In a live pipeline, inference runs once per frame.
• Input lag — The delay between an action and its visible result on screen. In a capture setup, the time for video to reach the PC and be processed adds to it.
• Input Translator — A configuration in ConsoleTuner’s Gtuner IV that maps controller buttons, keyboard keys and mouse movement to the Titan Two’s controller output. It loads into one of the Titan Two’s memory slots.
• Intersection over union (IoU) — The overlap area of two boxes divided by the total area they cover, from 0 (no overlap) to 1 (identical). It scores detections against ground truth and drives NMS.
• IRONAIM — A computer-vision module by IRONHIDE Software that runs inside the IRONICS launcher. Like every computer-vision module in IRONICS, it needs a capture card.
• IRONICS — IRONHIDE Software’s Windows desktop launcher for PS5, Xbox and PC. One app hosts IRONHIDE’s computer-vision modules, including IRONAIM, with built-in vision models and several modes, plus input plugins such as DigitalXIM and Virtual Titan, and translates controller, keyboard and mouse input for the target platform. It supports three output paths: Titan Two, Remote Play and Makcu. More on the IRONICS page.

K–N: Keypoint to NPU

• Keypoint — A specific point that a pose or landmark model locates, such as an elbow, knee or fingertip, stored as x and y coordinates with a confidence score.
• Latency budget — The share of the frame time each pipeline stage may use so the whole chain finishes within one frame.
• Makcu — A USB passthrough device and one of the three ways IRONICS sends controller output to the console. IRONICS drives it through the built-in Virtual Titan.
• mAP (mean average precision) — The standard accuracy metric for object detectors. “mAP 50–95” averages precision over IoU thresholds from 0.50 to 0.95; higher is better.
• Model — A trained neural network: an architecture plus learned weights. Common file formats are .pt for PyTorch and .onnx for ONNX.
• NMS (non-maximum suppression) — A post-processing step that removes duplicate detections: among overlapping boxes of the same class, it keeps the one with the highest confidence. Some newer detectors, such as YOLO26, can skip it with an end-to-end head.
• NPU (neural processing unit) — A processor block dedicated to neural-network inference, found in recent laptop chips such as those in Copilot+ PCs. Windows ML can reach NPUs through vendor execution providers.

O–R: OCR to Remote Play

• OCR (optical character recognition) — Turning pixels of text into machine-readable text. In gaming it reads scoreboards, timers and menus for stats logs or text-to-speech accessibility tools.
• ONNX — Open Neural Network Exchange, an open model format that defines a common set of operators and a common file format, so a model trained in one framework runs in many runtimes.
• ONNX Runtime — Microsoft’s cross-platform engine for running ONNX models on CPUs, GPUs and NPUs through execution providers.
• Opset — The operator set version an ONNX model was exported with. The runtime and execution provider must support it; the DirectML provider supports up to opset 20.
• P5 General — A PS5 adapter for the Titan Two. On PS5, the Titan Two additionally needs an adapter of this kind to work.
• Passthrough — A capture card’s HDMI output that forwards the console signal to your TV or monitor in real time, so you play on the display while the PC receives a copy.
• Polling rate — How often a USB input device reports its state, in hertz. At 1,000 Hz the device reports every millisecond; at 125 Hz, every 8 ms.
• Pose estimation — The task of locating keypoints on a body or object, for example to drive an avatar overlay from a streamer’s webcam.
• Post-processing — The steps after inference that turn raw model outputs into results: confidence filtering, NMS and scaling coordinates back to the original frame.
• Preprocessing — The steps before inference that turn a frame into model input: resizing (often to 640 × 640), color conversion, scaling pixel values to 0–1 and packing them into a tensor.
• Quantization (INT8) — Storing weights and activations as 8-bit integers instead of floats. It shrinks the model and can speed up inference, but needs calibration data to limit accuracy loss.
• Remote Play — PS Remote Play and Xbox Remote Play stream a console’s video to a PC over the network and send controller input back. In IRONICS, the Virtual Titan drives Remote Play with no adapter hardware.

S–Z: Segmentation to YOLO

• Segmentation — The task of labeling pixels. Instance segmentation outputs a mask per object; semantic segmentation assigns a class to every pixel.
• Tensor — A multi-dimensional array of numbers. One 640 × 640 RGB image in a batch of one is a tensor of shape 1 × 3 × 640 × 640.
• TensorRT — NVIDIA’s inference optimizer and runtime. ONNX Runtime’s TensorRT execution provider builds a GPU-specific engine on first run, which is why engine caching matters.
• Titan Two — A programmable controller adapter by ConsoleTuner, configured with Gtuner IV. It supports keyboard and mouse and is one of the three output paths in IRONICS. See what the Titan Two is.
• Tracking — Following objects across frames by giving each one a persistent ID, usually by matching new detections to positions predicted by a Kalman filter. ByteTrack and BoT-SORT are common trackers.
• Training — Teaching a model from labeled data by adjusting its weights to reduce errors. Fine-tuning starts from a pretrained model and needs far less data.
• USB OTG (On-The-Go) — A USB feature that lets a portable device act as a host, for example a phone that accepts a USB controller or keyboard.
• UVC (USB Video Class) — The USB standard for video devices. UVC-compliant capture cards and webcams work with the driver built into Windows, usbvideo.sys, with no vendor driver required.
• Virtual Titan — The software controller built into IRONICS. It drives PS Remote Play or Xbox Remote Play on the PC without adapter hardware, and it also drives a Makcu.
• VRR (variable refresh rate) — An HDMI 2.1 gaming feature that lets the display update when a new frame is ready instead of on a fixed schedule, reducing lag, judder and tearing.
• Windows ML — Microsoft’s Windows-maintained copy of ONNX Runtime with the same APIs. It can install hardware-specific execution providers through Windows Update and now receives the feature work DirectML no longer gets.
• XIM-style adapter — A hardware adapter that connects a keyboard and mouse to a console and presents them as a controller, converting mouse movement into analog-stick input. The name comes from the XIM product line.
• YOLO — “You Only Look Once,” a family of real-time object detectors that predict all boxes in one pass. Current Ultralytics versions include YOLO11 and YOLO26.

Frequently asked questions

Sources

• ConsoleTuner — Titan Two
• ConsoleTuner — Gtuner IV
• ConsoleTuner — Titan Two keyboard and mouse Input Translators
• Elgato — Game Capture HD60 X PlayStation 5 setup
• Elgato — Game Capture HD60 X technical specifications
• HDMI Licensing Administrator — Variable Refresh Rate
• PlayStation Support — How to change PS5 console resolution
• Xbox — Xbox Series X
• Microsoft Learn — USB Video Class driver overview
• USB-IF — Human Interface Devices (HID)
• USB-IF — USB On-The-Go
• PlayStation — PS Remote Play
• Xbox Support — How to set up remote play
• Microsoft Learn — Introduction to DirectML
• Microsoft Learn — What is Windows ML?
• ONNX — Open Neural Network Exchange
• ONNX Runtime — DirectML execution provider
• ONNX Runtime — TensorRT execution provider
• Ultralytics — YOLO26 model documentation
• Ultralytics — Multi-object tracking