This paper presents a method for learning object categories from just their names using the raw output of image search engines like Google. The approach introduces a new model called TSI-pLSA, which extends probabilistic Latent Semantic Analysis (pLSA) to incorporate spatial information in a translation and scale-invariant manner. The model is designed to handle the high intra-class variability and large proportion of unrelated images returned by search engines. The method is evaluated on standard test sets, showing performance competitive with existing methods trained on manually prepared datasets. The paper discusses the challenges of object category recognition in computer vision, where current methods require manually collected training data. The proposed approach leverages the vast amount of images available through image search engines, which can be noisy and contain many irrelevant images. However, the model is able to learn effective classifiers from such data, enabling automatic learning of visual categories. The paper describes the TSI-pLSA model, which incorporates spatial information by modeling the position of object centroids within images. This allows the model to handle variations in pose and scale. The model is trained using a combination of image search results and is tested on various datasets, including the Caltech and PASCAL datasets. The results show that the TSI-pLSA model performs well, particularly in handling pose variations and localizing objects in images. The paper also discusses the use of different approaches, including pLSA, ABS-pLSA, and TSI-pLSA, and evaluates their performance on various tasks. The results indicate that TSI-pLSA outperforms the other models in certain scenarios, particularly when dealing with pose variations. The model is also shown to be effective in improving the quality of image search results by re-ranking images based on learned topics. The paper concludes that the proposed method is effective in learning object categories from image search results, achieving performance competitive with existing methods that require manually prepared datasets. The approach has the potential to significantly reduce the need for manual data collection in object recognition tasks.This paper presents a method for learning object categories from just their names using the raw output of image search engines like Google. The approach introduces a new model called TSI-pLSA, which extends probabilistic Latent Semantic Analysis (pLSA) to incorporate spatial information in a translation and scale-invariant manner. The model is designed to handle the high intra-class variability and large proportion of unrelated images returned by search engines. The method is evaluated on standard test sets, showing performance competitive with existing methods trained on manually prepared datasets. The paper discusses the challenges of object category recognition in computer vision, where current methods require manually collected training data. The proposed approach leverages the vast amount of images available through image search engines, which can be noisy and contain many irrelevant images. However, the model is able to learn effective classifiers from such data, enabling automatic learning of visual categories. The paper describes the TSI-pLSA model, which incorporates spatial information by modeling the position of object centroids within images. This allows the model to handle variations in pose and scale. The model is trained using a combination of image search results and is tested on various datasets, including the Caltech and PASCAL datasets. The results show that the TSI-pLSA model performs well, particularly in handling pose variations and localizing objects in images. The paper also discusses the use of different approaches, including pLSA, ABS-pLSA, and TSI-pLSA, and evaluates their performance on various tasks. The results indicate that TSI-pLSA outperforms the other models in certain scenarios, particularly when dealing with pose variations. The model is also shown to be effective in improving the quality of image search results by re-ranking images based on learned topics. The paper concludes that the proposed method is effective in learning object categories from image search results, achieving performance competitive with existing methods that require manually prepared datasets. The approach has the potential to significantly reduce the need for manual data collection in object recognition tasks.