Instance search based on weakly supervised feature learning

Lin, Jie; Yu, Zhan; Zhao, Wan‐Lei

doi:10.1016/j.neucom.2019.11.029

Cited by 9 publications

(21 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The instances that the trained model could detect are restricted to a limited number of object categories. PCL*+SPN [20] extracts features from the object detection framework trained with image-level features. Despite leveraging on weakly supervised networks, similar retrieval performance as FCIS+XD is reported in [20].…”

Section: Related Work 21 Instance Searchmentioning

confidence: 99%

“…PCL*+SPN [20] extracts features from the object detection framework trained with image-level features. Despite leveraging on weakly supervised networks, similar retrieval performance as FCIS+XD is reported in [20]. The pitfall of this approach is that the network requires extra training stages and its discriminativeness towards the unknown categories is undermined due to the extra training.…”

Section: Related Work 21 Instance Searchmentioning

confidence: 99%

“…To this end, three groups of comparative studies are presented. The performance of DASR is studied in comparison to R-MAC [37], CroW [16], CAM [15], BLCF [24], BLCF-SalGAN [24], Regional Attention [17], DeepVision [32], FCIS+XD [43] and PCL*+SPN [20] in the instance search task. Note that the comparison is based on instance-level matching.…”

Section: Experiments 41 Datasets and Experimental Setupmentioning

confidence: 99%

“…Given the detected salient region, the feature map from each convolutional layer could be used to derive the feature descriptor. Nevertheless, it has been widely witnessed that the search performance varies across diferent layers [2,20,43]. For this reason, ablation analysis is conducted to seek for the best suitable layer for instance search.…”

Section: Ablation Studymentioning

confidence: 99%

“…In contrast, the second group of approaches ine-tune the pre-trained CNNs with extra training examples in the COCO dataset. The only approach that falls into this group is PCL*+SPN [20]. The last group of approaches are DeepVision [32] and FCIS+XD [43], which leverage Faster R-CNN [31] and FCN respectively to extract features.…”

Section: Comparison Tomentioning

confidence: 99%

See 4 more Smart Citations

Deeply Activated Salient Region for Instance Search

Xiao

Zhao

Jie

et al. 2022

ACM Trans. Multimedia Comput. Commun. Appl.

Self Cite

View full text Add to dashboard Cite

The performance of instance search relies heavily on the ability to locate and describe a wide variety of object instances in a video/image collection. Due to the lack of a proper mechanism for locating instances and deriving feature representation, instance search is generally only effective when the instances are from known object categories. In this paper, a simple but effective instance-level feature representation approach is presented. Different from the existing approaches, the issues of class-agnostic instance localization and distinctive feature representation are considered. The former is achieved by detecting salient instance regions from an image by a layer-wise back-propagation process. The back-propagation starts from the last convolution layer of a pre-trained CNNs that is originally used for classification. The back-propagation proceeds layer-by-layer until it reaches the input layer. This allows the salient instance regions in the input image from both known and unknown categories to be activated. Each activated salient region covers the full, or more usually, a major range of an instance. The distinctive feature representation is produced by average-pooling on the feature map of a certain layer with the detected instance region. Experiments show that this kind of feature representation demonstrates considerably better performance than most of the existing approaches.

show abstract

Section: Related Work 21 Instance Searchmentioning

confidence: 99%

Section: Related Work 21 Instance Searchmentioning

confidence: 99%