This paper improves recent methods for large-scale image search by analyzing the bag-of-features (BOF) representation in the context of approximate nearest neighbor search. The authors propose two enhancements: Hamming embedding (HE) and weak geometric consistency (WGC). HE provides binary signatures that refine visual words, while WGC filters descriptors inconsistent in terms of angle and scale. These methods are integrated into the inverted file system and are efficient for large datasets. Experiments on a million-image dataset show significant improvements in performance due to HE and WGC. HE enhances matching by using Hamming distances, while WGC improves accuracy by enforcing geometric consistency. The paper also discusses the voting interpretation of BOF, the limitations of quantization-based approaches, and the integration of HE and WGC into the inverted file. The results demonstrate that HE and WGC significantly improve image search accuracy, especially for large datasets. The methods are efficient and do not increase runtime significantly. The paper concludes that HE and WGC are effective for large-scale image search, with HE providing binary signatures and WGC enforcing geometric consistency. The results show that combining HE and WGC improves performance, and re-ranking with full geometric verification further enhances accuracy. The paper also evaluates the impact of different parameters and datasets, showing that HE and WGC outperform traditional BOF methods.This paper improves recent methods for large-scale image search by analyzing the bag-of-features (BOF) representation in the context of approximate nearest neighbor search. The authors propose two enhancements: Hamming embedding (HE) and weak geometric consistency (WGC). HE provides binary signatures that refine visual words, while WGC filters descriptors inconsistent in terms of angle and scale. These methods are integrated into the inverted file system and are efficient for large datasets. Experiments on a million-image dataset show significant improvements in performance due to HE and WGC. HE enhances matching by using Hamming distances, while WGC improves accuracy by enforcing geometric consistency. The paper also discusses the voting interpretation of BOF, the limitations of quantization-based approaches, and the integration of HE and WGC into the inverted file. The results demonstrate that HE and WGC significantly improve image search accuracy, especially for large datasets. The methods are efficient and do not increase runtime significantly. The paper concludes that HE and WGC are effective for large-scale image search, with HE providing binary signatures and WGC enforcing geometric consistency. The results show that combining HE and WGC improves performance, and re-ranking with full geometric verification further enhances accuracy. The paper also evaluates the impact of different parameters and datasets, showing that HE and WGC outperform traditional BOF methods.