Object Recognition
How it works
The input image passes through a feature-extraction network (backbone — a CNN or vision transformer) that builds hierarchical representations. In classification, a classification head maps features to a probability distribution over classes. In detection, additional heads predict bounding boxes and classes for many objects at once — two-stage approaches (e.g. R-CNN) first generate region proposals, while single-stage approaches (e.g. YOLO, SSD) detect in a single network pass. Overlapping detections are filtered with non-maximum suppression (NMS). Networks are trained on labeled datasets via backpropagation.
Problem solved
How to automatically determine which objects appear in an image and where they are, robustly to changes in lighting, scale, pose, occlusion, and background. Classical methods based on hand-crafted features (e.g. SIFT, HOG) generalized poorly; deep learning addressed this by learning representations directly from data.
Components
A network (CNN or vision transformer) that turns the image into feature maps consumed by downstream heads.
Maps feature representations to a probability distribution over a defined set of classes.
Outputs box coordinates and class labels, enabling detection of multiple objects in a single image.
A post-processing step that keeps the most confident box among overlapping detections of the same object using an IoU threshold.
Implementation
Small objects occupy few pixels and are hard to detect, especially after backbone downsampling.
Partially occluded or crowded objects lead to missed or merged detections.
A model trained in one domain loses accuracy under different lighting, cameras, or environments — critical in robotics.
Background dominating over objects skews training of single-stage detectors.
Poorly chosen thresholds cause false duplicates or removal of correct detections.
Evolution
A large-scale, hierarchical image dataset enabling training and benchmarking of object recognition.
A deep CNN dramatically lowered ImageNet classification error, starting the dominance of deep networks in computer vision.
Combining region proposals with CNN features substantially improved object detection accuracy.
Single-pass detection framed as regression, enabling real-time operation.
Residual connections enabled training of very deep networks, further raising image recognition accuracy.
Hardware requirements
Dense convolution and attention operations in CNNs/ViTs are strongly accelerated by GPU tensor cores.
TPUs efficiently run the large matrix multiplications of training and inference for vision networks.
Lightweight models can run on CPU, but at lower throughput than on GPU.