Open-Set Recognition: A Good Closed-Set Classifier is All You Need
Authors: Sagar Vaze, Kai Han, Andrea Vedaldi, Andrew Zisserman
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we first demonstrate that the ability of a classifier to make the none-of-above decision is highly correlated with its accuracy on the closed-set classes. We find that this relationship holds across loss objectives and architectures, and further demonstrate the trend both on the standard OSR benchmarks as well as on a large-scale Image Net evaluation. Second, we use this correlation to boost the performance of the maximum softmax probability OSR baseline by improving its closed-set accuracy, and with this strong baseline achieve state-of-the-art on a number of OSR benchmarks. |
| Researcher Affiliation | Academia | Sagar Vaze Kai Han Andrea Vedaldi Andrew Zisserman Visual Geometry Group, University of Oxford The University of Hong Kong {sagar,vedaldi,az}@robots.ox.ac.uk kaihanx@hku.hk |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | Yes | Code available at: https://github.com/sgvaze/osr_closed_set_all_you_need. |
| Open Datasets | Yes | MNIST (Le Cun et al., 2010), SVHN (Netzer et al., 2011), CIFAR10 (Krizhevsky, 2009): These are ten-class datasets... Tiny Image Net (Le & Yang, 2015)... Image Net-21K-P (Ridnik et al., 2021)... Caltech-UCSD Birds (CUB) (Wah et al., 2011), Stanford Cars (Krause et al., 2013) FGVC-Aircraft (Maji et al., 2013). |
| Dataset Splits | Yes | In all cases, the model is trained on a subset of classes, while other classes are reserved as unseen for evaluation. MNIST (Le Cun et al., 2010), SVHN (Netzer et al., 2011), CIFAR10 (Krizhevsky, 2009): ...training on six classes, while using the other four classes for testing (|C| = 6; |U| = 4). CIFAR + N ... training on four classes from CIFAR10, while using N classes from CIFAR100 for evaluation, where N denotes either 10 or 50 classes (|C| = 4; |U| {10, 50}). Tiny Image Net ... 20 classes used for training and 180 as unknown (|C| = 20; |U| = 180). ... For both Easy and Hard splits, we have |C| = 1000 and |U| = 1000. ... We select the Rand Augment and label smoothing hyper-parameters by maximizing closed-set accuracy on a validation set (randomly sampling 20% of the training set). |
| Hardware Specification | Yes | We train all models for 600 epochs with a batch size of 128, training models on a single NVIDIA Titan X GPU. ... the memory intensive nature of the method meant we could only fit a batch size of 2 on a 12GB GPU. We attempted to scale it up for the FGVC datasets, fitting a batch size of 16 across 4 24GB GPUs, with training taking a week. |
| Software Dependencies | No | No specific version numbers for key software components (e.g., Python, PyTorch, CUDA libraries) were found. The paper mentions using a 'Res Net50 model pre-trained with the cross-entropy loss on Image Net-1K from (Wightman, 2019)' which points to 'Pytorch image models' but does not specify their own direct dependencies with versions. |
| Experiment Setup | Yes | We train all models for 600 epochs with a batch size of 128. ... We use an initial learning rate of 0.1 for all datasets except Tiny Image Net, for which we use 0.01. We train with a cosine annealed learning rate, restarting the learning rate to the initial value at epochs 200 and 400. Furthermore, we warm up the learning rate by linearly increasing it from 0 to the initial value at epoch 20. ... We use Rand Augment for all experiments... We follow a similar procedure for the label smoothing value s, though we find the optimal value to be s = 0 for all datasets except Tiny Image Net, where it helps significantly at s = 0.9. |