Research map / Architectures and efficiency / CNN design
Making Convolutional Networks Recurrent for Visual Sequence Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Transform pre-trained CNN layers into recurrent layers | VGG16 with GRU · 300VW test | AUC: 0.722 -> 0.772 | Helped |
| Transform pre-trained CNN layers into recurrent layers | VGG16 with LSTM · 300VW test | AUC: 0.718 -> 0.769 | Helped |
| Use two recurrent layers rather than one (ablation) | Traditional LSTM on VGG16 · 300VW test | AUC: 0.718 -> 0.671 | Hurt |
| Use two recurrent layers rather than one (ablation) | Traditional GRU on VGG16 · 300VW test | AUC: 0.722 -> 0.698 | Hurt |
| Use PreGRU at fc6 and fc7 instead of only fc6 (ablation) | VGG16 · 300VW test | AUC: 0.755 -> 0.772 | Helped |
| Share one input-to-hidden state across GRU gates (ablation) | PreGRU on VGG16 · 300VW test | AUC: 0.772 -> 0.762 | No clear effect |
| Transform pre-trained CNN layers into recurrent layers | C3D with LSTM · NVGesture test | classification accuracy: 81.3% -> 81.9% | Helped |
| Use PreRNN-SIH rather than gate-dependent PreRNN (ablation) | C3D with LSTM · NVGesture test | classification accuracy: 82.7% -> 84.2% | Helped |
| Use PreRNN-SIH rather than gate-dependent PreRNN (ablation) | C3D with GRU · NVGesture test | classification accuracy: 83.1% -> 83.8% | Mixed |
| Use recurrent CNN layers instead of no RNN | C3D on NVGesture Color + Depth · NVGesture test | classification accuracy: 81.0% -> 85.0% | Helped |
| Transform the pre-trained ResNet layer into a recurrent layer | ResNet50 with LSTM · UCF101 first split, Color | classification accuracy: 83.4% -> 85.3% | Helped |
| Transform the pre-trained ResNet layer into a recurrent layer | ResNet50 with GRU · UCF101 first split, Color + Flow | classification accuracy: 92.2% -> 93.7% | Helped |
Closest papers by meaning
- Deep RNN Framework for Visual Sequential Applications (CVPR 2019 · similarity 0.90)
- Convolutional Sequence to Sequence Learning (ICML 2017 · similarity 0.89)
- Task-Driven Convolutional Recurrent Models of the Visual System (NeurIPS 2018 · similarity 0.88)
- Learning Compact Recurrent Neural Networks With Block-Term Tensor Decomposition (CVPR 2018 · similarity 0.88)
- Structural-RNN: Deep Learning on Spatio-Temporal Graphs (CVPR 2016 · similarity 0.88)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · similarity 0.87)
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.