This paper evaluates several learning methods for recognizing generic visual categories with invariance to pose, lighting, and surrounding clutter. A large dataset of 194,400 images was collected, consisting of 50 uniform-colored toys under 36 angles, 9 azimuths, and 6 lighting conditions. The toys were categorized into five generic classes: four-legged animals, human figures, airplanes, trucks, and cars. Five instances of each category were used for training, and the other five for testing. Low-resolution grayscale images with varying degrees of variability and clutter were used for training and testing. The methods tested included Nearest Neighbor, Support Vector Machines (SVM), and Convolutional Networks (CNN) operating on raw pixels or PCA-derived features. Test error rates for unseen object instances on uniform backgrounds were around 13% for SVM and 7% for CNN. On a segmentation/recognition task with highly cluttered images, SVM proved impractical, while CNN yielded 14% error. A real-time system was implemented that can detect and classify objects in natural scenes at around 10 frames per second. The paper describes the NORB dataset, a large image dataset comprising 97,200 stereo image pairs of 50 objects belonging to five generic categories under various conditions. The dataset was designed to reflect real imaging situations by preserving natural variabilities and eliminating irrelevant clues. The dataset was used to evaluate the performance of various image classification methods, including K-Nearest Neighbors, SVM, and CNN. The results showed that CNNs outperformed SVMs and K-NN in terms of accuracy and efficiency, especially in handling complex transformations such as pose and lighting. CNNs achieved error rates below 7% on the normalized-uniform dataset and were more robust to variations in the input data. The paper also highlights the importance of trainable local feature extractors for robust and invariant recognition. The paper concludes that the NORB dataset provides a benchmark for evaluating learning-based approaches to invariant object recognition. It emphasizes the importance of trainable local feature extractors for robust and invariant recognition. The paper also suggests that future work will use trainable classifiers that incorporate explicit models of image formation and geometry.This paper evaluates several learning methods for recognizing generic visual categories with invariance to pose, lighting, and surrounding clutter. A large dataset of 194,400 images was collected, consisting of 50 uniform-colored toys under 36 angles, 9 azimuths, and 6 lighting conditions. The toys were categorized into five generic classes: four-legged animals, human figures, airplanes, trucks, and cars. Five instances of each category were used for training, and the other five for testing. Low-resolution grayscale images with varying degrees of variability and clutter were used for training and testing. The methods tested included Nearest Neighbor, Support Vector Machines (SVM), and Convolutional Networks (CNN) operating on raw pixels or PCA-derived features. Test error rates for unseen object instances on uniform backgrounds were around 13% for SVM and 7% for CNN. On a segmentation/recognition task with highly cluttered images, SVM proved impractical, while CNN yielded 14% error. A real-time system was implemented that can detect and classify objects in natural scenes at around 10 frames per second. The paper describes the NORB dataset, a large image dataset comprising 97,200 stereo image pairs of 50 objects belonging to five generic categories under various conditions. The dataset was designed to reflect real imaging situations by preserving natural variabilities and eliminating irrelevant clues. The dataset was used to evaluate the performance of various image classification methods, including K-Nearest Neighbors, SVM, and CNN. The results showed that CNNs outperformed SVMs and K-NN in terms of accuracy and efficiency, especially in handling complex transformations such as pose and lighting. CNNs achieved error rates below 7% on the normalized-uniform dataset and were more robust to variations in the input data. The paper also highlights the importance of trainable local feature extractors for robust and invariant recognition. The paper concludes that the NORB dataset provides a benchmark for evaluating learning-based approaches to invariant object recognition. It emphasizes the importance of trainable local feature extractors for robust and invariant recognition. The paper also suggests that future work will use trainable classifiers that incorporate explicit models of image formation and geometry.