{"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/190409460","title":"Data-Driven Neuron Allocation for Scale Aggregation Networks","arxiv_id":"1904.09460","date":"2019-04-20","proceeding":"CVPR 2019 6","authors":["Yi Li","Zhanghui Kuang","Yimin Chen","Wayne Zhang"],"abstract":"Successful visual recognition networks benefit from aggregating information\nspanning from a wide range of scales. Previous research has investigated\ninformation fusion of connected layers or multiple branches in a block, seeking\nto strengthen the power of multi-scale representations. Despite their great\nsuccesses, existing practices often allocate the neurons for each scale\nmanually, and keep the same ratio in all aggregation blocks of an entire\nnetwork, rendering suboptimal performance. In this paper, we propose to learn\nthe neuron allocation for aggregating multi-scale information in different\nbuilding blocks of a deep network. The most informative output neurons in each\nblock are preserved while others are discarded, and thus neurons for multiple\nscales are competitively and adaptively allocated. Our scale aggregation\nnetwork (ScaleNet) is constructed by repeating a scale aggregation (SA) block\nthat concatenates feature maps at a wide range of scales. Feature maps for each\nscale are generated by a stack of downsampling, convolution and upsampling\noperations. The data-driven neuron allocation and SA block achieve strong\nrepresentational power at the cost of considerably low computational\ncomplexity. The proposed ScaleNet, by replacing all 3x3 convolutions in ResNet\nwith our SA blocks, achieves better performance than ResNet and its outstanding\nvariants like ResNeXt and SE-ResNet, in the same computational complexity. On\nImageNet classification, ScaleNets absolutely reduce the top-1 error rate of\nResNets by 1.12 (101 layers) and 1.82 (50 layers). On COCO object detection,\nScaleNets absolutely improve the mmAP with backbone of ResNets by 3.6 (101\nlayers) and 4.6 (50 layers) on Faster RCNN, respectively. Code and models are\nreleased at https://github.com/Eli-YiLi/ScaleNet.","url_abs":"http://arxiv.org/abs/1904.09460v1","url_pdf":"http://arxiv.org/pdf/1904.09460v1.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":"190409460","repo_url":"https://github.com/Eli-YiLi/ScaleNet","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":null}],"tasks":[{"task_slug":"image-classification","task_name":"Image Classification"},{"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":"average-pooling","method_name":"Average Pooling"},{"method_slug":"batch-normalization","method_name":"Batch Normalization"},{"method_slug":"bottleneck-residual-block","method_name":"Bottleneck Residual Block"},{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"faster-r-cnn","method_name":"Faster R-CNN"},{"method_slug":"global-average-pooling","method_name":"Global Average Pooling"},{"method_slug":"grouped-convolution","method_name":"Grouped Convolution"},{"method_slug":"kaiming-initialization","method_name":"Kaiming Initialization"},{"method_slug":"max-pooling","method_name":"Max Pooling"},{"method_slug":"rpn","method_name":"RPN"},{"method_slug":"randomhorizontalflip","method_name":"Random Horizontal Flip"},{"method_slug":"random-resized-crop","method_name":"Random Resized Crop"},{"method_slug":"relu","method_name":"ReLU"},{"method_slug":"resnext","method_name":"ResNeXt"},{"method_slug":"resnext-block","method_name":"ResNeXt Block"},{"method_slug":"residual-block","method_name":"Residual Block"},{"method_slug":"residual-connection","method_name":"Residual Connection"},{"method_slug":"roipool","method_name":"RoIPool"},{"method_slug":"sgd-with-momentum","method_name":"SGD with Momentum"},{"method_slug":"scale-aggregation-block","method_name":"Scale Aggregation Block"},{"method_slug":"scalenet","method_name":"ScaleNet"},{"method_slug":"softmax","method_name":"Softmax"},{"method_slug":"step-decay","method_name":"Step Decay"},{"method_slug":"weight-decay","method_name":"Weight Decay"}],"datasets_introduced":[],"methods_introduced":[{"slug":"scale-aggregation-block","name":"Scale Aggregation Block","full_name":"Scale Aggregation Block"},{"slug":"scalenet","name":"ScaleNet","full_name":"ScaleNet"}],"results":[{"leaderboard":"/sota/image-classification-on-imagenet","task":"Image Classification","dataset":"ImageNet","model":"ScaleNet-152","rank_in_archive_order":763,"of":1060,"metrics":{"GFLOPs":"11.2","Top 1 Accuracy":"79.38%"},"uses_additional_data":false},{"leaderboard":"/sota/image-classification-on-imagenet","task":"Image Classification","dataset":"ImageNet","model":"ScaleNet-101","rank_in_archive_order":787,"of":1060,"metrics":{"GFLOPs":"7.5","Top 1 Accuracy":"79.03%"},"uses_additional_data":false},{"leaderboard":"/sota/image-classification-on-imagenet","task":"Image Classification","dataset":"ImageNet","model":"ScaleNet-50","rank_in_archive_order":863,"of":1060,"metrics":{"GFLOPs":"3.8","Top 1 Accuracy":"77.8%"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1904.09460","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}