{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/object-detection-from-scratch-with-deep","title":"Object Detection from Scratch with Deep Supervision","arxiv_id":"1809.09294","date":"2018-09-25","proceeding":null,"authors":["Zhiqiang Shen","Zhuang Liu","Jianguo Li","Yu-Gang Jiang","Yurong Chen","xiangyang xue"],"abstract":"We propose Deeply Supervised Object Detectors (DSOD), an object detection\nframework that can be trained from scratch. Recent advances in object detection\nheavily depend on the off-the-shelf models pre-trained on large-scale\nclassification datasets like ImageNet and OpenImage. However, one problem is\nthat adopting pre-trained models from classification to detection task may\nincur learning bias due to the different objective function and diverse\ndistributions of object categories. Techniques like fine-tuning on detection\ntask could alleviate this issue to some extent but are still not fundamental.\nFurthermore, transferring these pre-trained models across discrepant domains\nwill be more difficult (e.g., from RGB to depth images). Thus, a better\nsolution to handle these critical problems is to train object detectors from\nscratch, which motivates our proposed method. Previous efforts on this\ndirection mainly failed by reasons of the limited training data and naive\nbackbone network structures for object detection. In DSOD, we contribute a set\nof design principles for learning object detectors from scratch. One of the key\nprinciples is the deep supervision, enabled by layer-wise dense connections in\nboth backbone networks and prediction layers, plays a critical role in learning\ngood detectors from scratch. After involving several other principles, we build\nour DSOD based on the single-shot detection framework (SSD). We evaluate our\nmethod on PASCAL VOC 2007, 2012 and COCO datasets. DSOD achieves consistently\nbetter results than the state-of-the-art methods with much more compact models.\nSpecifically, DSOD outperforms baseline method SSD on all three benchmarks,\nwhile requiring only 1/2 parameters. We also observe that DSOD can achieve\ncomparable/slightly better results than Mask RCNN + FPN (under similar input\nsize) with only 1/3 parameters, using no extra data or pre-trained models.","url_abs":"http://arxiv.org/abs/1809.09294v2","url_pdf":"http://arxiv.org/pdf/1809.09294v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"object-detection-from-scratch-with-deep","repo_url":"https://github.com/szq0214/DSOD","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"caffe2","reach":{"status":"ok","spdx":"NOASSERTION"}}],"tasks":[{"task_slug":"classification","task_name":"General Classification"},{"task_slug":"object","task_name":"Object"},{"task_slug":"object-detection","task_name":"Object Detection"},{"task_slug":"object-detection-1","task_name":"object-detection"}],"methods":[{"method_slug":"1x1-convolution","method_name":"1x1 Convolution"},{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"fpn","method_name":"FPN"},{"method_slug":"non-maximum-suppression","method_name":"Non Maximum Suppression"},{"method_slug":"ssd","method_name":"SSD"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1809.09294","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}