Live Camera Stream
Click below to grant camera access and start real-time zero-lag object detection.
Confidence Threshold
Filter out low-scoring bounding boxes
Verified Detections
Live active objects meeting threshold
Precision-Recall Curve (IoU = 0.50)
YOLOv7 Model Loss Curves & Convergence History
CIoU Bounding Box Localization, Objectness BCE, and Multi-Class Cross-Entropy
Measures bounding box overlap, center distance, and aspect ratio consistency against ground truth annotations. Starts high (~0.082) and drops to ~0.019.
Evaluates whether an anchor grid cell contains a real object or background noise. Uses Binary Cross-Entropy with focal scaling to suppress background false positives.
Measures category identification accuracy across 80 classes. Decreases to ~0.005 as the model learns distinctive visual features.
Neural vision models require square 640×640 inputs. Rather than squishing your camera feed, VisionStream applies letterbox padding bars so objects preserve their 100% true physical proportions.
Stream Geometry Properties Zero Distortion
Live Detected Objects Geometry
Live 640×640 Square Tensor Feed
Live Video PreviewGray letterbox bars show zero-padding sent to YOLOv7 to prevent distortion.
Standard webcams stream at 16:9 (1.78:1 widescreen) or 4:3 (1.33:1). Neural networks like YOLOv7 process strict uniform 1:1 square inputs (640×640).
Directly squishing a widescreen video into a square stretches people and cars unnaturally, reducing detection score by ~15%. Letterboxing adds padding bars with 0.00% distortion.
A tensor is a multi-dimensional array of numbers. YOLOv7 takes an input tensor of shape [1, 3, 640, 640] (Batch=1, RGB=3, 640×640 pixels) and outputs [1, 25200, 85] anchor detections.