(2.13 MB) Abstract
Insight-DCU participated in the instance search (INS), semantic indexing (SIN), and localization tasks (LOC) this year.
In the INS task we used deep convolutional network features trained on external data and the query data for this year to train our system. We submitted four runs, three based on convolutional network features, and one based on SIFT/BoW. F A insightdcu 1 was an automatic run using features from the last convolutional layer of a deep network with bag-of-words encoding and achieved 0.123 mAP. F A insightdcu 2 modied the previous run to use re-ranking based on an R-CNN model and achieved 0.111 mAP. I A insightdcu 3, our interactive run, achieved 0.269 mAP. Our SIFT-based run F A insightdcu 2 used weak geometric consistency to improve performance over the previous year to 0.187 mAP. Overall we found that using features from the convolutional layers improved performance over features from the fully connected layers used in previous years, and that weak geometric consistency improves performance for local feature ranking.
In the SIN task we again used convolutional network features, this time netuning a network pretrained on external data for the task. We submitted four runs, 2C D A insightdcu.15 1..4 varying the top-level learning algorithm and use of concept co-occurance. 2C D A insightdcu.15 1 used a linear SVM top-level learner, and achieved 0.63 mAP. Exploiting concept co-occurance improved the accuracy of our logistic regression run 2C D A insightdcu.15 3 from 0.058 mAP to 0.6 2C D A insightdcu.15 3.
Our LOC system used training data from IACC.1.B and features similar to our INS run, but using a VLAD encoding instead of a bag-of-words. Unfortunately there was problem with the run that we are still investigating.
Note: UPC and NII participated only in the INS task of this submission.
