Learning from the Mistakes of Others: Matching Errors in Cross-Dataset Learning

Adaptive SVM+: Learning with Privileged Information for Domain Adaptation

“…Although the improvement can be seen as marginal in the linear kernel case, note that all methods are marginally better than a regular SVM, since some interactions are very similar to some others (e.g., walking to, walking away from, walking with), which makes the accurate classification of such tasks very challenging. Adaptive SVM+ is more accurate in 32 out of the 60 interactions, SVM MMD [28] in 19, and the rest are attributed to SVM, SVM+ and Adaptive SVM. When RBF kernels are used, there is a 2.81% relative improvement.…”

“…Additionally, illustrations in the form of clip art are provided depicting the same 60 fine-grained categories in two different level settings: (i) category-level in which images and illustrations are collected independently, and (ii) instance-level in which 2-3 illustrations of the same interaction category are collected for a given image. We followed the same experimental procedure with the method of Sharmanska and Quadrianto [28] for the instance-level setting. They proposed a framework called SVM MMD to "learn from the mistakes of others" by minimizing the distribution mismatch between errors made in images and in privileged data (i.e., illustrations) using the Maximum Mean Discrepancy (MMD) criterion.…”

“…Features: Both real images and illustrations are represented by a 765-dimensional feature vector capturing human pose information, expressions, relation (from person 1 to person 2) and appearance and are provided with the dataset. As in [28] we pair each real image with a randomly selected (among the two or three) illustration per image. Clip art illustrations are used as a source domain and real images as a target domain.…”

“…The train/test split process is repeated 20 times and average results along with standard error across repeats are reported. Model Selection: Following the evaluation protocol of Sharmanska and Quadrianto [28], we select the parameter C from {10 0 , . .…”

“…Training: We trained 60 one-versus-rest binary classifiers to predict the interaction of interest against the rest of the interactions. Similar to [28], we trained Adaptive SVM+ using linear kernels and by sampling 25 positive vs 25 negative images. For testing, we use the remaining positive images balanced with negative samples.…”