Unveiling the Potential of Structure Preserving for Weakly Supervised Object Localization

Pan, Xingjia; Gao, Yingguo; Lin, Zhiwen; Tang, Fan; Dong, Weiming; Yuan, Haolei; Huang, Feiyue; Xu, Changsheng

doi:10.1109/cvpr46437.2021.01147

Cited by 62 publications

(33 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Following the previous methods SPA [13], we apply both bounding box and mask metrics to evaluate the performance of our BAS. For bounding box, following [13,20,25], we use three metrics for evaluation, including GT-known localization accuracy (GT-known Loc), Top-1 localization accuracy (Top-1 Loc), and Top-5 localization accuracy (Top-5 Loc). Specifically, GT-known Loc is correct when the intersection over union(IoU) between the ground-truth bounding box and the predicted bounding box is greater than 0.5.…”

Section: Methodsmentioning

confidence: 99%

“…the correlation of different pictures of the same class. Besides, SPA [13] uses post-processing to extract feature maps with structure-preserving. SLT [6] considers several similar classes as one class when generating classification loss and localization maps, which alleviates the problem of focusing on the most discriminative regions by increasing tolerance.…”

Section: Related Workmentioning

confidence: 99%

“…When L BAS is larger than 1, we mark it as 1, to ensure the stability of the initial training. In the inference phase, we use ten crop augmentations to get the final classification results following the settings in [6,13,28]. For localization, we replace the random crop Ours CAM [31] Image Ours CAM [31] Image CUB-200-2011 [19] ILSVRC [14] Figure 3.…”

Section: Implementation Detailsmentioning

confidence: 99%

“…Segmentation Quality. We compare the localization map with the ground-truth mask label using two metrics, Peak T and Peak IoU, following SPA [13]. As shown in Fig.…”

Section: Performance Analysismentioning

confidence: 99%

“…To alleviate the problem, some methods [4,11,16,24,27] erase the most discriminative regions during network learning, forcing the network to learn more object features relevant to localization. [13,28,29] are also based on CAM, which improves localization performance by establishing pixel-level spatial correlation. Additionally, some methods [6,10,20,25] suggest adopting a divide-andconquer strategy to accomplish classification and localization tasks separately to avoid conflicts.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Background Activation Suppression for Weakly Supervised Object Localization

Wu¹,

Zhai²,

Cao³

2021

Preprint

View full text Add to dashboard Cite

Weakly supervised object localization (WSOL) aims to localize the object region using only image-level labels as supervision. Recently a new paradigm has emerged by generating a foreground prediction map (FPM) to achieve the localization task. Existing FPM-based methods use crossentropy (CE) to evaluate the foreground prediction map and to guide the learning of generator. We argue for using activation value to achieve more efficient learning. It is based on the experimental observation that, for a trained network, CE converges to zero when the foreground mask covers only part of the object region. While activation value increases until the mask expands to the object boundary, which indicates that more object areas can be learned by using activation value. In this paper, we propose a Background Activation Suppression (BAS) method. Specifically, an Activation Map Constraint module (AMC) is designed to facilitate the learning of generator by suppressing the background activation values. Meanwhile, by using the foreground region guidance and the area constraint, BAS can learn the whole region of the object. Furthermore, in the inference phase, we consider the prediction maps of different categories together to obtain the final localization results. Extensive experiments show that BAS achieves significant and consistent improvement over the baseline methods on the CUB-200-2011 and ILSVRC datasets 1 .

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Implementation Detailsmentioning

confidence: 99%

“…Segmentation Quality. We compare the localization map with the ground-truth mask label using two metrics, Peak T and Peak IoU, following SPA [13]. As shown in Fig.…”

Section: Performance Analysismentioning

confidence: 99%