Research map / Detection and segmentation / CNN features
OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OverLoCK, a pure ConvNet backbone with top-down "overview-first, look-closely-next" decomposition (Base-Net, Overview-Net, Focus-Net) and a context-mixing dynamic convolution (ContMix), reaching 84.2% ImageNet top-1 for the Tiny variant.
Problem
Modern ConvNets use a pyramid structure with no explicit top-down semantic guidance, and existing convolutions struggle to combine long-range dependency modeling with strong local inductive biases at high resolution.
Approach
A branched backbone (Deep-stage Decomposition) where a lightweight Overview-Net produces a coarse global context prior that is fused into a Focus-Net, plus ContMix, a dynamic convolution whose spatially varying kernels are built from token-to-region-center affinities over the top-down context.
Key results
- OverLoCK-T reaches 84.2% ImageNet-1K top-1 at 5.5G FLOPs and 33M params, improving UniRepLKNet-T by 1% and VMamba-T by 1.6%.
- OverLoCK-S surpasses MogaNet-B by 1% APb with Mask R-CNN 1x schedule (49.4 vs 47.9 APb on the table).
- OverLoCK-T improves UniRepLKNet-T by 1.7% mIoU on ADE20K (50.3 vs 48.6 under UperNet 160K).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OverLoCK-T instead of MogaNet-S | Image classification models · ImageNet-1K val, 224×224 | Top-1 accuracy: 83.4% to 84.2% | Helped |
| Use OverLoCK-S instead of MogaNet-B | Image classification models · ImageNet-1K val, 224×224 | Top-1 accuracy: 84.3% to 84.8% | Helped |
| Use OverLoCK-S instead of MogaNet-B | Mask R-CNN · COCO, 1× schedule | APb: +1.5% | Helped |
| Use OverLoCK-S instead of PeLK-S | Cascade Mask R-CNN · COCO, 3× schedule | APb: +1.4% | Helped |
| Use OverLoCK-T instead of UniRepLKNet-T | UperNet · ADE20K | mIoU: +1.7% | Helped |
| Replace vanilla convolution with Dilated RepConv (ablation) | PlainNet · ImageNet-1K (120 epochs) | Top-1 accuracy: 76.3% to 76.6% | Helped |
| Add an SE Layer (ablation) | PlainNet with Dilated RepConv · ImageNet-1K (120 epochs) | Top-1 accuracy: 76.6% to 77.1% | Helped |
| Add a 3×3 DWConv (ablation) | PlainNet with SE · ImageNet-1K (120 epochs) | Top-1 accuracy: 77.1% to 78.0% | Helped |
| Replace vanilla FFN with ConvFFN (ablation) | PlainNet with local convolution · ImageNet-1K (120 epochs) | Top-1 accuracy: 78.0% to 78.5% | Helped |
| Use a recurrent top-down design (ablation) | Baseline network · ImageNet-1K (120 epochs) | Top-1 accuracy: 78.5% to 76.8% | Hurt |
| Apply DDS decomposition (ablation) | Baseline network · ImageNet-1K (120 epochs) | Top-1 accuracy: 78.5% to 79.0% | Helped |
| Remove Base-Net feature feed into Focus-Net (ablation) | DDS Model · ImageNet-1K (120 epochs) | Top-1 accuracy: 79.0% to 78.2% | Hurt |
Surprising findings
- A recurrent top-down variant (AbsViT-style) reduced accuracy (76.8% vs 76.6% baseline-free comparison at lower efficiency), suggesting recurrent designs are unsuitable for modern ConvNets.
- Feeding only the Overview-Net output into Focus-Net without Base-Net features lowered performance (78.2% vs 79.0%).
Limitations
- Gap between classification and dense prediction for ConvNets is motivated but limited gains are shown only against selected baselines.
- Ablations are run at 120 epochs on XT-scale models and a single downstream task (ADE20K segmentation).
Experimental setup
Datasets: ImageNet-1K COCO 2017 ADE20K
Benchmarks: ImageNet-1K COCO 2017 ADE20K
Metrics: Top-1 accuracy FLOPs parameters throughput APb APm mIoU effective receptive field
Baselines: ConvNeXt UniRepLKNet MogaNet VMamba BiFormer PeLK InternImage MaxViT Swin DyConv ODConv Involution
Builds on: ConvNeXt UniRepLKNet DCNv3 / InternImage Dilated RepConv AbsViT
Closest papers by meaning
- Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation (AAAI 2026 · similarity 0.85)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.85)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.85)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.85)
- Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.85)
- CanViT: Toward Active-Vision Foundation Models (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.