Research map / Detection and segmentation / CNN features
DeNet: Scalable Real-Time Object Detection With Directed Sparse Sampling
Lachlan Tychsen-Smith, Lars Petersson
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| raise corner loss weight lambda_t from 1 to 100 (ablation) | DeNet-34 · Pascal VOC 2007 val | mAP (%): 49.6 -> 72.8 | Improved |
| raise lambda_t further from 100 to 500 (ablation) | DeNet-34 · Pascal VOC 2007 val | mAP (%): 72.8 -> 70.6 | Hurt |
| raise bbox regression weight lambda_b from 1 to 10 at lambda_t=100 (ablation) | DeNet-34 · Pascal VOC 2007 val | mAP (%): 72.8 -> 72.1 | Hurt |
| add skip layers (ablation) | DeNet-34 · MSCOCO test-dev2015 | AP@IoU=0.5:0.95 (%): 29.4 -> 29.5 | Improved |
| add skip layers (ablation) | DeNet-34 · MSCOCO test-dev2015 | AP@0.5 (%): 46.2 -> 47.9 | Improved |
| add skip layers (ablation) | DeNet-101 · MSCOCO test-dev2015 | AP@IoU=0.5:0.95 (%): 31.9 -> 32.3 | Improved |
| use wide variant (128x128 corner maps, N=48) on top of skip (ablation) | DeNet-34 · MSCOCO test-dev2015 | AP@IoU=0.5:0.95 (%): 29.5 -> 30.0 | Improved |
| use wide variant on top of skip (ablation) | DeNet-101 · MSCOCO test-dev2015 | AP@IoU=0.5:0.95 (%): 32.3 -> 33.8 | Improved |
| use wide variant on top of skip (evaluation rate cost) (ablation) | DeNet-101 · MSCOCO test-dev2015 | evaluation rate (Hz): 33 Hz -> 17 Hz | Hurt |
| add skip layers (ablation) | DeNet-34 · Pascal VOC 2007 test | mAP (%): 75.3% -> 75.9% | Improved |
| add skip layers (ablation) | DeNet-101 · Pascal VOC 2007 test | mAP (%): 77.0% -> 77.1% | No clear effect |
| increase training/testing sample boxes from 64 to 1024 (ablation) | DeNet-34 · Pascal VOC 2007 | coverage (%) and eval rate (Hz): coverage 81 -> 97, rate 96 Hz -> 55 Hz | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- R-FCN — R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016) · better on 1 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Sparse R-CNN: An End-to-End Framework for Object Detection (TPAMI 2023 · similarity 0.92)
- Sparse R-CNN: End-to-End Object Detection With Learnable Proposals (CVPR 2021 · similarity 0.92)
- Region-Based Convolutional Networks for Accurate Object Detection and Segmentation (TPAMI 2016 · similarity 0.91)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.91)
- Deep Learning for Generic Object Detection: A Survey (IJCV 2019 · similarity 0.90)
- Learning Rich Features at High-Speed for Single-Shot Object Detection (ICCV 2019 · similarity 0.90)
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.
