ICDAR 2015 refers to the Robust Reading Competition held at the 13th International Conference on Document Analysis and Recognition (ICDAR) in 2015. The competition focused on scene text detection and recognition, challenging participants to locate and read text in natural images. It was part of a series of competitions that have driven progress in the field of Computer vision and document analysis, providing standardized benchmarks for evaluating algorithms.
The competition comprised several tasks, most notably text localization (detecting bounding boxes around text regions) and text recognition (transcribing the detected text into a machine-readable string). The primary dataset used was the ICDAR 2015 Incidental Scene Text dataset, which consisted of 1,500 images captured with a Google Glass device in a busy street environment. These images were intentionally captured in an 'incidental' manner, meaning they were not carefully framed, leading to challenges such as motion blur, low resolution, and arbitrary orientations of text.
Dataset and Evaluation
The ICDAR 2015 dataset included 1,000 training images and 500 test images. Ground truth annotations provided word-level bounding boxes and transcriptions. Evaluation for detection used the standard PASCAL VOC criterion, where a detection was considered correct if the intersection-over-union (IoU) with a ground truth box exceeded 0.5. For recognition, the metric was the edit distance between the predicted and ground truth strings. The competition also included a combined task that required both detection and recognition, evaluated using the end-to-end word spotting metric.
Results and Impact
The winning entry for the text localization task was submitted by a team from the University of Toronto, achieving an F-measure of approximately 0.72. For the end-to-end recognition task, the top performer was a team from the Chinese Academy of Sciences, reaching an F-measure of about 0.68. These results were modest compared to later years, reflecting the difficulty of the incidental scene text problem. The competition highlighted the limitations of existing methods in handling arbitrary orientations and varying lighting conditions, spurring subsequent research in deep learning-based approaches.
Legacy and Follow-up
ICDAR 2015 became a widely used benchmark in the research community. Many subsequent papers on scene text detection and recognition, particularly those employing Deep learning and Neural network architectures, reported results on this dataset. The competition's emphasis on incidental text influenced later editions, such as ICDAR 2017 and ICDAR 2019, which introduced even more challenging datasets. The 2015 benchmark remains a standard reference for evaluating the robustness of text reading systems in real-world scenarios.
Relation to Broader AI Research
The techniques developed for ICDAR 2015, such as region proposal networks and sequence-to-sequence models for recognition, intersected with broader advances in Artificial intelligence and Machine learning. The competition served as a practical testbed for algorithms that later found applications in autonomous driving, augmented reality, and document digitization. It also contributed to the development of Data Augmentation strategies, as participants employed synthetic text generation and image transformations to improve generalization.
See Also
- ICDAR 2013
- Scene text recognition
- Object detection
- Convolutional neural network