Dynamic Pseudo-Label Generation for Weakly Supervised Object Detection in Remote Sensing Images

Abstract

In recent years, fully supervised object detection methods in remote sensing images with good performance have been developed. However, this approach requires a large number of instance-level annotated samples that are relatively expensive to acquire. Therefore, weakly supervised learning using only image-level annotations has attracted much attention. Most of the weakly supervised object detection methods are based on multi-instance learning methods, and their performance depends on the process of scoring the candidate region proposals during training. In this process, the use of only image-level labels for supervision usually cannot obtain optimal results due to the lack of location information of the object. To address the above problem, a dynamic sample pseudo-label generation framework is proposed to generate pseudo-labels for each proposal without additional annotations. First, we propose the pseudo-label generation algorithm (PLG) to generate the category labels of the proposal by using the localization information of the object. Specifically, we propose to use the pixel average of the object’s localization map in the proposal as the proposal category confidence and calculate the pseudo-label by comparing the proposal category confidence with the preset threshold. In addition, an effective adaptive threshold selection strategy is designed to eliminate the effect of different category shape differences in computing sample pseudo-labels. Comparative experiments on the NWPU VHR-10 dataset demonstrate that our method can significantly improve the detection performance compared to existing methods.

Proposals Generation for Weakly Supervised Object Detection in Artwork Images

“…Many studies combine CAMs [ 37 ] or Weakly Supervised Semantic Segmentation (WSSS) [ 38 ] to achieve better WSOD performances. The authors of [ 39 , 40 , 41 ] leverage the power of CAMs as segmentation proposals, [ 42 , 43 , 44 , 45 ] introduce a collaboration loop between the segmentation and detection branches, [ 46 ] proposes a cascaded convolutional neural network and [ 47 ] exploit segmentation properties, i.e., purity and completeness, to harvest tight boxes that take into account the surrounding context. Still, the actual methods cannot fully exploit CAMs as bounding box generators and require the use of external domain-dependent proposals or hybrid-annotated data.…”

A Comprehensive Review of Emerging Technologies: Machine Learning and UAV in Crop Management

“…Main objective of ML is, to classify the data based on models developed, make predictions for future and outcomes based on these models [12]. ML is broadly classified into prediction and classification [20]. Figure 1.…”

“…Focused research must be conducted on fusing complementary sensors and sensing abilities, simple manipulators must be developed and enhanced path optimization, navigation and path guidance algorithms suitable for dynamic climate conditions must be developed [15]. A great number of fully supervised object detection methods based on CNN have been proposed for better performance [20]. But these deep learning-based methods seeking large number of samples with instance level labels is laborious and time taking process for interpreting images with many targets [20].…”

“…A great number of fully supervised object detection methods based on CNN have been proposed for better performance [20]. But these deep learning-based methods seeking large number of samples with instance level labels is laborious and time taking process for interpreting images with many targets [20]. Therefore WSOD, Weakly supervised object detection is proposed that require only image level labels along with location information of the object [20].…”

Supervised and Weakly Supervised Deep Learning for Segmentation and Counting of Cotton Bolls Using Proximal Imagery

“…When the lower-level labels are available, many approaches are available to localize and detect object instances [ 20 ] that leverage two prominent paradigms: multiple-instance learning (MIL) [ 21 ] and class feature activation maps (CAMs) [ 22 ]. MIL is based on learning object instances from positive (one or more instances present) or negative (no instances) bins of data samples, and this approach has been used widely for weakly supervised object detection [ 23 , 24 ]. CAMs use intermediate feature maps from the classifier’s activation layers along with pseudo-label generation, and the method has gained popularity due to its versatility for object detection as well as for instance segmentation [ 25 , 26 , 27 ].…”