Research map / Detection and segmentation / CNN features
Training-Time-Friendly Network for Real-Time Object Detection
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 |
|---|---|---|---|
| encode multiple regression samples in a Gaussian sub-area instead of only the object-center sample | TTFNet versus CenterNet · COCO 5k-val | AP: 20.8 -> 28.1 | Helped |
| use logarithmic box-area scaling with normalized sample weights rather than equal weighting (ablation) | TTFNet-53 · COCO 5k-val | AP: 27.2 -> 32.0 | Helped |
| add Gaussian probabilities to the normalized logarithmic regression weights (ablation) | TTFNet-53 · COCO 5k-val | AP: 32.0 -> 32.9 | Helped |
| increase rectangular regression sub-area from β=0.01 to β=0.3 (ablation) | TTFNet-53 · COCO 5k-val | AP: 27.2 -> 32.5 | Helped |
| increase rectangular regression sub-area from β=0.3 to β=0.9 (ablation) | TTFNet-53 · COCO 5k-val | AP: 32.5 -> 29.5 | Hurt |
| use class-aware regression at β=0.9 (ablation) | TTFNet-53 · COCO 5k-val | AP: 29.5 -> 22.6 | Hurt |
| consider box aspect ratio in the rectangular sub-area (ablation) | TTFNet-53, β=0.3 · COCO 5k-val | AP: 32.2 -> 32.5 | Helped |
| use Gaussian rather than rectangular regression sub-area with aspect ratio (ablation) | TTFNet-53 · COCO 5k-val | AP: 32.5 -> 32.9 | Helped |
| add shortcut connections to backbone stages (ablation) | TTFNet-53 · COCO 5k-val | AP: 29.0 -> 32.0 | Helped |
| increase shortcut layers to 3/3/1 for stages 2/3/4 (ablation) | TTFNet-53 · COCO 5k-val | AP: 32.9 -> 33.2 | Helped |
| increase regression kernel size and encoded samples while raising the learning rate (ablation) | TTFNet-53, 2x schedule · COCO 5k-val | AP: 33.1 -> 35.3 | Helped |
| train from scratch with a 10x schedule rather than use pretrained weights with a 2x schedule (ablation) | TTFNet, ResNet-18 · MS COCO 2017 val | AP: 28.1 -> 30.3 | Helped |
Closest papers by meaning
- Optimizing Video Object Detection via a Scale-Time Lattice (CVPR 2018 · similarity 0.88)
- Mimicking Very Efficient Network for Object Detection (CVPR 2017 · similarity 0.88)
- ScratchDet: Training Single-Shot Object Detectors From Scratch (CVPR 2019 · similarity 0.87)
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.87)
- Rethinking Scale-Aware Temporal Encoding for Event-based Object Detection (NeurIPS 2025 · similarity 0.87)
- SCNN: A General Distribution Based Statistical Convolutional Neural Network with Application to Video Object Detection (AAAI 2019 · 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.