{"id":{"repo_id":"malta","oai_identifier":"oai:www.um.edu.mt:123456789/64106"},"canonical_url":"https://search.dev.ndltd.org/etd/malta/oai:www.um.edu.mt:123456789/64106","repository":{"repo_id":"malta","name":"University of Malta","base_url":"https://www.um.edu.mt/library/oar/oai/request"},"display":{"title":"A study on the effect of target object size in object detection","abstract":"Recent years have seen an impressive increase in the efficiency and accuracy of object detection models through the use of Region-based Convolutional Neural Networks (RCNN). Studies have been carried out to improve object detection models. However, the detection of small objects still poses numerous challenges for the said models. Small object detection is considered as one of the biggest challenges in object detection for several reasons. One in particular is due to the resizing of the feature maps within the pooling stage resulting in the loss of the small target object’s features. Moreover, accuracy diminishes as these networks struggle to distinguish between foreground and complex backgrounds such as rough terrain. An experiment was designed to investigate a selection of deep neural networks and analyse the effects that the distance between the capturing device and the target object. The set up requires a custom dataset, containing object instances that are identifiable by the selected models, each with an object instance at a varying distance. The COCO dataset was selected as the training and benchmarking dataset in this experiment. Following the selection of a benchmarking dataset, the custom dataset used to test for the effect of target object size was obtained. The custom dataset was prepared in a controlled environment set up with pre-established distance measurements and multiple small objects placed in a non-complex background for detection. By varying the range of distances, size of the object instance decreases, producing a lower amount of features. The selection of state-of-the-art object detection models include YOLO, EfficientDet and Detectron2 which are all pre-trained on the COCO dataset. To evaluate each model, a series of baseline tests were initially carried out to ensure that each model could correctly identify the object instance that is being used as a ground truth. Once the baselines were established, the custom dataset was applied and the results were extracted. Results show that EfficientDet produces an average FBeta measure of 67.2%, a 20.7% increase from the second best performing model for the custom dataset when measuring over all distances. However, it was interesting to note that accuracy falls off once the object instance exceeds five or more meters, with the object instance going either undetected or incorrectly labelled.","abstract_html":"Recent years have seen an impressive increase in the efficiency and accuracy of object detection models through the use of Region-based Convolutional Neural Networks (RCNN). Studies have been carried out to improve object detection models. However, the detection of small objects still poses numerous challenges for the said models. Small object detection is considered as one of the biggest challenges in object detection for several reasons. One in particular is due to the resizing of the feature maps within the pooling stage resulting in the loss of the small target object’s features. Moreover, accuracy diminishes as these networks struggle to distinguish between foreground and complex backgrounds such as rough terrain. An experiment was designed to investigate a selection of deep neural networks and analyse the effects that the distance between the capturing device and the target object. The set up requires a custom dataset, containing object instances that are identifiable by the selected models, each with an object instance at a varying distance. The COCO dataset was selected as the training and benchmarking dataset in this experiment. Following the selection of a benchmarking dataset, the custom dataset used to test for the effect of target object size was obtained. The custom dataset was prepared in a controlled environment set up with pre-established distance measurements and multiple small objects placed in a non-complex background for detection. By varying the range of distances, size of the object instance decreases, producing a lower amount of features. The selection of state-of-the-art object detection models include YOLO, EfficientDet and Detectron2 which are all pre-trained on the COCO dataset. To evaluate each model, a series of baseline tests were initially carried out to ensure that each model could correctly identify the object instance that is being used as a ground truth. Once the baselines were established, the custom dataset was applied and the results were extracted. Results show that EfficientDet produces an average FBeta measure of 67.2%, a 20.7% increase from the second best performing model for the custom dataset when measuring over all distances. However, it was interesting to note that accuracy falls off once the object instance exceeds five or more meters, with the object instance going either undetected or incorrectly labelled.","abstract_has_math":false,"creators":[],"institution":"University of Malta","degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2020,"date_issued":"2020","date_published":"2020","updated_at":"2026-07-27T20:12:39Z","subjects":["Pattern recognition systems","Neural networks (Computer science)","Data sets"],"languages":["en"],"rights":["info:eu-repo/semantics/restrictedAccess"],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://www.um.edu.mt/library/oar/handle/123456789/64106","outbound_label":"Repository record","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2020-11-18T12:01:16Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2020-11-18T12:01:16Z"]},{"key":"dc:date.issued","label":"Date","values":["2020"]},{"key":"dc:publisher.department","label":"Dc Publisher Department","values":["Faculty of Information and Communication Technology. Department of Artificial Intelligence"]},{"key":"dc:publisher.institution","label":"Dc Publisher Institution","values":["University of Malta"]},{"key":"dc:type","label":"Dc Type","values":["bachelorThesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Pattern recognition systems","Neural networks (Computer science)","Data sets"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.iso","label":"Language (ISO)","values":["en"]},{"key":"dc:rights","label":"Dc Rights","values":["info:eu-repo/semantics/restrictedAccess"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://www.um.edu.mt/library/oar/handle/123456789/64106"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["B.SC.ICT(HONS)ARTIFICIAL INTELLIGENCE"]},{"key":"dc:description.abstract","label":"Abstract","values":["Recent years have seen an impressive increase in the efficiency and accuracy of object detection models through the use of Region-based Convolutional Neural Networks (RCNN). Studies have been carried out to improve object detection models. However, the detection of small objects still poses numerous challenges for the said models. Small object detection is considered as one of the biggest challenges in object detection for several reasons. One in particular is due to the resizing of the feature maps within the pooling stage resulting in the loss of the small target object’s features. Moreover, accuracy diminishes as these networks struggle to distinguish between foreground and complex backgrounds such as rough terrain. An experiment was designed to investigate a selection of deep neural networks and analyse the effects that the distance between the capturing device and the target object. The set up requires a custom dataset, containing object instances that are identifiable by the selected models, each with an object instance at a varying distance. The COCO dataset was selected as the training and benchmarking dataset in this experiment. Following the selection of a benchmarking dataset, the custom dataset used to test for the effect of target object size was obtained. The custom dataset was prepared in a controlled environment set up with pre-established distance measurements and multiple small objects placed in a non-complex background for detection. By varying the range of distances, size of the object instance decreases, producing a lower amount of features. The selection of state-of-the-art object detection models include YOLO, EfficientDet and Detectron2 which are all pre-trained on the COCO dataset. To evaluate each model, a series of baseline tests were initially carried out to ensure that each model could correctly identify the object instance that is being used as a ground truth. Once the baselines were established, the custom dataset was applied and the results were extracted. Results show that EfficientDet produces an average FBeta measure of 67.2%, a 20.7% increase from the second best performing model for the custom dataset when measuring over all distances. However, it was interesting to note that accuracy falls off once the object instance exceeds five or more meters, with the object instance going either undetected or incorrectly labelled."]},{"key":"dc:title","label":"Title","values":["A study on the effect of target object size in object detection"]}]}],"canonical_facts":{"dc:date.accessioned":["2020-11-18T12:01:16Z"],"dc:date.available":["2020-11-18T12:01:16Z"],"dc:date.issued":["2020"],"dc:description":["B.SC.ICT(HONS)ARTIFICIAL INTELLIGENCE"],"dc:description.abstract":["Recent years have seen an impressive increase in the efficiency and accuracy of object detection models through the use of Region-based Convolutional Neural Networks (RCNN). Studies have been carried out to improve object detection models. However, the detection of small objects still poses numerous challenges for the said models. Small object detection is considered as one of the biggest challenges in object detection for several reasons. One in particular is due to the resizing of the feature maps within the pooling stage resulting in the loss of the small target object’s features. Moreover, accuracy diminishes as these networks struggle to distinguish between foreground and complex backgrounds such as rough terrain. An experiment was designed to investigate a selection of deep neural networks and analyse the effects that the distance between the capturing device and the target object. The set up requires a custom dataset, containing object instances that are identifiable by the selected models, each with an object instance at a varying distance. The COCO dataset was selected as the training and benchmarking dataset in this experiment. Following the selection of a benchmarking dataset, the custom dataset used to test for the effect of target object size was obtained. The custom dataset was prepared in a controlled environment set up with pre-established distance measurements and multiple small objects placed in a non-complex background for detection. By varying the range of distances, size of the object instance decreases, producing a lower amount of features. The selection of state-of-the-art object detection models include YOLO, EfficientDet and Detectron2 which are all pre-trained on the COCO dataset. To evaluate each model, a series of baseline tests were initially carried out to ensure that each model could correctly identify the object instance that is being used as a ground truth. Once the baselines were established, the custom dataset was applied and the results were extracted. Results show that EfficientDet produces an average FBeta measure of 67.2%, a 20.7% increase from the second best performing model for the custom dataset when measuring over all distances. However, it was interesting to note that accuracy falls off once the object instance exceeds five or more meters, with the object instance going either undetected or incorrectly labelled."],"dc:identifier.uri":["https://www.um.edu.mt/library/oar/handle/123456789/64106"],"dc:language.iso":["en"],"dc:publisher.department":["Faculty of Information and Communication Technology. Department of Artificial Intelligence"],"dc:publisher.institution":["University of Malta"],"dc:rights":["info:eu-repo/semantics/restrictedAccess"],"dc:subject":["Pattern recognition systems","Neural networks (Computer science)","Data sets"],"dc:title":["A study on the effect of target object size in object detection"],"dc:type":["bachelorThesis"]},"updated_at":"2026-07-27T20:12:39Z"}