Towards Writer Retrieval for Historical Datasets

Evaluating learned feature aggregators for writer retrieval

“…4) plots, we can observe very similar behavior across all objective functions, suggesting a theoretical limitation rather than a limitation in the objective. It is worth noting that the results achieved here are not far off from the existing stateof-the-art: For example, [1] reports a precision of between 87.6% and 88.5% depending on which ResNet-backbone is chosen. 1 This is very similar to our results (see Table 1) of ≈ 83% precision.…”

“…It is worth noting that the results achieved here are not far off from the existing stateof-the-art: For example, [1] reports a precision of between 87.6% and 88.5% depending on which ResNet-backbone is chosen. 1 This is very similar to our results (see Table 1) of ≈ 83% precision. However, the transformer-based method is still consistently weaker than the simple NetRVLAD [1] produced by existing models.…”

“…In general, the results of our end-to-end models still loses notably against the model build by [1]. We hypothesize that this is due to the larger hypothesis set present in end-to-end models: If one trains the feature aggregator jointly with the rest of the model, not only does one have additional freedom in how to aggregate information, but also what information is being extracted in the first place.…”

“…These patches are then processed separately inside some unsupervised computer vision model, such as a ResNet [12], and external mechanisms, such as Vector of Locally Aggregated Descriptors (VLAD) encodings, which are employed to merge the individual patches into a combined embedding. Recent works mostly focus on improving the feature extraction stage (see, for example, [1,5]), but not the information aggregation phase. These methods rely on hand-crafted features [13][14][15][16][17][18][19][20][21] like SIFT descriptors [15,16,18,21].…”

“…To our knowledge, VICReg is the strictest general-purpose self-supervised learning technique. One would presumably need additional constraints on the neural network to close the gap to e. g., [1]…”