Learning visual biases from human imagination
Authors: Carl Vondrick, Hamed Pirsiavash, Aude Oliva, Antonio Torralba
NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments are surprising, and suggest that classifiers from the human visual system can be transferred into a machine with some success. |
| Researcher Affiliation | Academia | Carl Vondrick Hamed Pirsiavash Aude Oliva Antonio Torralba Massachusetts Institute of Technology University of Maryland, Baltimore County {vondrick,oliva,torralba}@mit.edu hpirsiav@umbc.edu |
| Pseudocode | No | The paper describes mathematical formulations but does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code (specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described in this paper. |
| Open Datasets | Yes | We quantify their performance on object classification in realworld images using the PASCAL VOC 2011 dataset [13], evaluating against the validation set. We trained an SVM classifier with CNN features to recognize cars on Caltech 101 [14], but we tested it on object classification with PASCAL VOC 2011. Furthermore, standard computer vision datasets often suffer from dataset biases that harm cross dataset generalization performance [32, 28]. Since the template we estimate is biased by the human visual system and not datasets (there is no dataset), we believe our approach may help cross dataset generalization. We trained an SVM classifier with CNN features to recognize cars on Caltech 101 [14], but we tested it on object classification with PASCAL VOC 2011. Fig.10a suggest that, by constraining the SVM to be close to the human bias for car, we are able to improve the generalization performance of our classifiers, sometimes over 5% AP. We then tried the reverse experiment in Fig.10b: we trained on PASCAL VOC 2011, but tested on Caltech 101. While PASCAL VOC provides a much better sample of the visual world, the orientation priors still help generalization performance when there is little training data available. |
| Dataset Splits | Yes | We quantify their performance on object classification in realworld images using the PASCAL VOC 2011 dataset [13], evaluating against the validation set. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions software like HOGgles and MOSEK, but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We wish to add the constraint that the SVM hyperplane w must be at most cos 1(θ) degrees away from the bias template c: ... In our experiments, we found 30 to be reasonable. While this angle is not very restrictive in low dimensions, it becomes much more restrictive as the number of dimensions increases [21]. |