Research map / Detection and segmentation / CNN features
Learn to Predict Sets Using Feed-Forward Neural Networks
Hamid Rezatofighi, Tianyu Zhu, Roman Kaskman, Farbod Motlagh, Javen Qinfeng Shi, Anton Milan, Daniel Cremers, Laura Leal-Taixé, Ian D. Reid
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| train classifier jointly with cardinality on a shared backbone (JDS) (ablation) | Set-DS (BCE-SftMx) · PASCAL VOC 2007 multi-label classification | C-F1: 76.2 -> 82.2 | Improved |
| train classifier jointly with cardinality on a shared backbone (JDS) (ablation) | Set-DS (BCE-SftMx) · PASCAL VOC 2007 multi-label classification | I-F1: 83.0 -> 92.5 | Improved |
| train classifier jointly with cardinality on a shared backbone (JDS) (ablation) | Set-DS (BCE-Sftmx) · MS COCO multi-label classification | O-F1: 69.6 -> 70.7 | Improved |
| train classifier jointly with cardinality on a shared backbone (JDS) (ablation) | Set-DS (BCE-Sftmx) · MS COCO multi-label classification | I-F1: 74.6 -> 75.6 | Improved |
| replace negative binomial cardinality loss with softmax categorical loss (ablation) | Set-DS (BCE) · PASCAL VOC 2007 multi-label classification | C-F1: 75.8 -> 76.2 | Improved |
| replace negative binomial cardinality loss with softmax categorical loss (ablation) | Set-DS (BCE) · MS COCO multi-label classification | C-F1: 64.6 -> 64.7 | No clear effect |
| learn cardinality jointly with the classifier (ablation) | Set network · PASCAL VOC 2007 multi-label classification | mean absolute cardinality error: 0.33 ± 0.53 -> 0.31 ± 0.54 | No clear effect |
| use set formulation instead of Faster R-CNN with the same ResNet101 and L1 loss | Faster R-CNN · Cropped MOTChallenge pedestrian detection | mAP: 0.68 -> 0.85 | Improved |
| use set formulation instead of sequential LSTM-Hungarian detector | LSTM-Hungarian · Cropped MOTChallenge pedestrian detection | mAP: 0.61 -> 0.85 | Improved |
| add GIoU to the L1 bounding box loss (ablation) | Set (Ours) ResNet101 · Cropped MOTChallenge pedestrian detection | MR: 0.25 -> 0.24 | No clear effect |
| add predicted cardinality at inference to DETR-style set detector (ablation) | DETR · Cropped MOTChallenge pedestrian detection | F1: 0.86 -> 0.87 | No clear effect |
| add predicted cardinality at inference to DETR-style set detector (ablation) | DETR · PASCAL VOC 2007 detection | F1: 0.71 -> 0.72 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Faster R-CNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · better on 1 of 1 shared result
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
- DeepSetNet: Predicting Sets With Deep Neural Networks (ICCV 2017 · similarity 0.95)
- Joint Learning of Set Cardinality and State Distribution (AAAI 2018 · similarity 0.93)
- Set Prediction in the Latent Space (NeurIPS 2021 · similarity 0.89)
- Deep Set Prediction Networks (NeurIPS 2019 · similarity 0.89)
- Towards Open Set Deep Networks (CVPR 2016 · similarity 0.87)
- Multiset-Equivariant Set Prediction with Approximate Implicit Differentiation (ICLR 2022 · similarity 0.86)
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.
