Density Map Guided Object Detection in Aerial Images

Li, Changlin; Yang, Taojiannan; Zhu, Sijie; Chen, Chen; Guan, Shanyue

doi:10.1109/cvprw50498.2020.00103

Cited by 189 publications

(103 citation statements)

References 21 publications

Supporting

Mentioning

103

Contrasting

Order By: Relevance

“…First, object detection needs to be conducted on certain areas excluding the massive backgrounds with very little chance of target objects being present [14,29]. Secondly, because objects taken in the aerial images are very small in pixels and distributed sparsely and non-uniformly, e.g., pedestrians or vehicles, object detectors can show improvement by re-executing object detection especially in zoom-out images whose target objects are densely crowded [16,30].…”

Section: Region Proposal Methods For Aerial Imagesmentioning

confidence: 99%

“…Given in a patch image, Pang et al [29] extracted a feature vector on the patch image via a lightweight residual network called Tiny-Net and a classifier takes the feature vector for the binary objectness prediction. For the latter purpose, a Density Map guided-detection Network called DMNet was presented in [30], which estimates a density map for a given input aerial image and crops connected regions based on the estimated density map. To be specific, DMNet obtains a density mask by applying a density threshold to the estimated density map and uses connected component algorithm to form the cropping connected regions.…”

Section: Region Proposal Methods For Aerial Imagesmentioning

confidence: 99%

See 1 more Smart Citation

Hierarchical Multi-Label Object Detection Framework for Remote Sensing Images

Shin

Kim

et al. 2020

Remote Sensing

View full text Add to dashboard Cite

Detecting objects such as aircraft and ships is a fundamental research area in remote sensing analytics. Owing to the prosperity and development of CNNs, many previous methodologies have been proposed for object detection within remote sensing images. Despite the advance, using the object detection datasets with a more complex structure, i.e., datasets with hierarchically multi-labeled objects, is limited to the existing detection models. Especially in remote sensing images, since objects are obtained from bird’s-eye view, the objects are captured with restricted visual features and not always guaranteed to be labeled up to fine categories. We propose a hierarchical multi-label object detection framework applicable to hierarchically partial-annotated datasets. In the framework, an object detection pipeline called Decoupled Hierarchical Classification Refinement (DHCR) fuses the results of two networks: (1) an object detection network with multiple classifiers, and (2) a hierarchical sibling classification network for supporting hierarchical multi-label classification. Our framework additionally introduces a region proposal method for efficient detection on vain areas of the remote sensing images, called clustering-guided cropping strategy. Thorough experiments validate the effectiveness of our framework on our own object detection datasets constructed with remote sensing images from WorldView-3 and SkySat satellites. Under our proposed framework, DHCR-based detections significantly improve the performance of respective baseline models and we achieve state-of-the-art results on the datasets.

show abstract

Section: Region Proposal Methods For Aerial Imagesmentioning

confidence: 99%

Section: Region Proposal Methods For Aerial Imagesmentioning

confidence: 99%

Hierarchical Multi-Label Object Detection Framework for Remote Sensing Images

Shin

Kim

et al. 2020

Remote Sensing

View full text Add to dashboard Cite

show abstract

“…The principle is to distinguish real targets and background interference by comparing the differences of targets in several consecutive frames of images before and after, so as to count birds in the continuous images shot by infrared camera. In literature [16], Density Map method was proposed to guide the detector to detect the vehicles in the aerial images. The basic idea is to firstly use the Density Map to calculate the density of the vehicles in the images, then, the image was segmented into several small pieces of different sizes according to the density map of the vehicles, and each piece of the image was detected separately with different anchors and intensity.…”

Section: B Aerial Image Detectionmentioning

confidence: 99%

“…Firstly, the cascade classifier is applied to identify the eyes and mouth of a human face; then, the physical feature relationship between the eyes and mouth of a person and the face is used to identify and detect the entire masked face. In the context of urban autonomous driving system, Only useful for objects of large scales in simple background [1], [3][4][5], [24] Feature fusion+ attention mechanism Extract features of very small objects(2×2 pixels) Only useful for objects in simple background such as sky and sea [2], [6] Density map Detect the small and high density objects Not useful for single object detection [16] Extra information Has the ability of reasoning and predict the occluded parts Detection on traditional visible datasets [17][18][19] Ours (secondary transfer learning + HNEM)…”

Section: Occlusion Detectionmentioning

confidence: 99%

See 1 more Smart Citation

Weak and Occluded Vehicle Detection in Complex Infrared Environment Based on Improved YOLOv4

Zhang

et al. 2021

IEEE Access

View full text Add to dashboard Cite

Infrared small target detection is still a challenge in the field of object detection. At present, although there are many related research achievements, it surely needs further improvement. This paper introduced a new application of severely occluded vehicle detection in the complex wild background of weak infrared camera aerial images, in which more than 50% area of the vehicles are occluded. We used YOLOv4 as the detection model. By applying secondary transfer learning from visible dataset to infrared dataset, the model could gain a good average precision (AP). Firstly, we trained the model in the UCAS_AOD visible dataset, then, we transferred it to the VIVID visible dataset, finally we transferred the model to the VIVID infrared dataset for a second training. Meanwhile, added the hard negative example mining block to the YOLOv4 model, which could depress the disturbance of complex background thus further decrease the false detecting rate. Through experiments the average precision improved from90.34% to 91.92%, the F1 score improved from 87.5% to 87.98%, which demonstrated that the proposed algorithm generated satisfactory and competitive vehicle detection results. INDEX TERMS Infrared aerial image, occlusion, vehicle detection, hard negative example mining, YOLOv4.

show abstract