Resisting Large Data Variations via Introspective Transformation Network

Zhao, Yunhan; Tian, Ye; Fowlkes, Charless C.; Shen, Wei; Yuille, Alan

doi:10.1109/wacv45572.2020.9093340

Cited by 2 publications

(5 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Qualitative results in Figure 4 shows the progression of examples generated by S. As training progresses, our approach generates increasingly hard examples in a variety of modes. Improvement in Accuracy: In Table 1, we compare our approach with recent methods [55,32,13,25,18] that generate synthetic data to improve accuracy on AffNIST data. For the result in Table 1, we use 55000, 5000, 10000 split for training, validation and testing as in [55] along with the same classifier architecture.…”

Section: Methodsmentioning

confidence: 99%

“…Improvement in Accuracy: In Table 1, we compare our approach with recent methods [55,32,13,25,18] that generate synthetic data to improve accuracy on AffNIST data. For the result in Table 1, we use 55000, 5000, 10000 split for training, validation and testing as in [55] along with the same classifier architecture. We outperform hard negative…”

Section: Methodsmentioning

confidence: 99%

“…Unlike our approach, these methods do not generate new examples. Recently, [18,55] proposed data augmentation for generating transformations that yields additional pseudo-negative training examples. In contrast, we generate hard positive examples.…”

Section: Related Workmentioning

confidence: 99%

“…generation approaches [55,25,18] by achieving a low error rate of 0.99%. Please find more details in the supplementary.…”

Section: Hardness Of Synthetic Data Easy Examples Hard Examplesmentioning

confidence: 99%

“…We attach the pytorch model dump for the synthesizer below. The acronyms used in the pytorch model dump are described in Table . For our experiments in Table . 1, we use the following architecture from [55], where Swish denotes the swish activation function from [33]. image dimension.…”

Section: Supplementary Material: Learning To Generate Synthetic Data ...mentioning

confidence: 99%

See 4 more Smart Citations

Learning to Generate Synthetic Data via Compositing

Tripathi

Chandra

Agrawal

et al. 2019

2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

104

View full text Add to dashboard Cite

We present TERSE, a task-aware approach to synthetic data generation. Our framework employs a trainable synthesizer network that is optimized to produce meaningful training samples by assessing the strengths and weaknesses of a 'target' network. The synthesizer and target networks are trained in an adversarial manner wherein each network is updated with a goal to outdo the other. Additionally, we ensure the synthesizer generates realistic data by pairing it with a discriminator trained on real-world images. Further, to make the target classifier invariant to blending artefacts, we introduce these artefacts to background regions of the training images so the target does not over-fit to them.We demonstrate the efficacy of our approach by applying it to different target networks including a classification network on AffNIST, and two object detection networks (SSD, Faster-RCNN) on different datasets. On the AffNIST benchmark, our approach is able to surpass the baseline results with just half the training examples. On the VOC person detection benchmark, we show improvements of up to 2.7% as a result of our data augmentation. Similarly on the GMU detection benchmark, we report a performance boost of 3.5% in mAP over the baseline method, outperforming the previous state of the art approaches by up to 7.5% on specific categories.

show abstract

Section: Methodsmentioning

confidence: 99%