ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models

Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Danny Gutfreund, Joshua Tenenbaum, Boris Katz

Abstract: We collect a large real-world test set, ObjectNet, for object recognition with controls where object backgrounds, rotations, and imaging viewpoints are random. Most scientific experiments have controls, confounds which are removed from the data, to ensure that subjects cannot perform a task by exploiting trivial correlations in the data. Historically, large machine learning and computer vision datasets have lacked such controls. This has resulted in models that perform better on datasets than in new applications and must be fine-tuned for most new datasets. When tested on ObjectNet, object detectors show a ~45% drop in performance, with respect to their performance on other benchmarks, when biases are removed. Reinserting biases, by subsetting the dataset along the axes which have controls, recovers most of this performance drop. ObjectNet is more robust to fine-tuning because of controls with only small performance increases. We develop a highly automated platform that enables gathering datasets with controls by crowdsourcing image capturing and annotation. ObjectNet is comparable in size to the ImageNet test set (40K vs 50K images),and by design does not come paired with a training set in order to encourage generalization. While we focus on object recognition, using automated tools data with controls can be gathered at scale throughout machine learning to generate datasets that exercise models in new ways providing valuable feedback to researchers. This work opens up new avenues for research in generalizable, robust, and more human-like computer vision and in creating datasets where results are diagnostic of expected real-world performance.

Code Link: https://objectnet.dev/code.html

CMT Num: 5037