Answers are grounded in this paper and cite their sources.
Convolutional neural networks and Transformer-based architectures have achieved remarkable progress in image classification and downstream vision tasks. However, multi-stage downsampling, while enhancing representation, often causes small-object information loss and limited receptive fields that hinder detection and segmentation. To address these issues, we propose DiamondBlock, a plug-and-play architectural module positioned at the front of backbone networks. DiamondBlock performs multiple rounds of upsampling on the original image, followed by symmetric downsampling and skip connections for multi-scale feature fusion. This design enhances fine-grained and large-scale representations with architecture-dependent parameter overhead but nontrivial computational overhead, since early-stage upsampling increases spatial resolution and thus FLOPs/latency depending on the number of DiamondBlock stages. Additionally, inspired by human memory mechanisms, we introduce a training paradigm based on the Ebbinghaus forgetting curve. Instead of purely random sampling, we design a periodic sample revisit strategy that simulates human review behavior to reinforce memory retention, requiring no architectural or optimizer changes. Extensive experiments on ImageNet-1K, COCO, and ADE20K validate the effectiveness of our approach. We explicitly report the accuracy-efficiency trade-off using FLOPs and throughput/latency, and recommend cost-aware configurations for different budgets. Integrating DiamondBlock into ViT backbones yields up to a 5.6% Top-1 accuracy improvement on ImageNet-1K. Moreover, under a step-matched training budget (i.e., keeping the total number of optimizer updates fixed), the Ebbinghaus-based training provides consistent but modest Top-1 gains when applied on top of DiamondBlockequipped backbones (typically around ∼0.1 percentage-point, up to +0.17 percentagepoint), improving generalization without architectural or optimizer changes. Overall, DiamondBlock combined with memory-driven training provides consistent improvements across classification, detection, and segmentation tasks.
No takes yet. Share an insight, caveat, or question.
Ning et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: