Synapse
⌘+K
Synapse
PulseExploreClubsResearchersJournals
Instagram
HomeClubsExplore
October 1, 2017

Focal Loss for Dense Object Detection

View Full Paper
Ask AI
Bookmark
Share

Authors

TLTsung-Yi LinNational Taipei University of TechnologyPGPriya GoyalDayanand Medical College & HospitalRGRoss GirshickAllen Institute

Discussion

Loading...

Member takes

Overview

Randomized trial demonstrates improved accuracy in dense object detection, suggesting novel approaches to loss functions can enhance performance.

Key Points

  • This research aims to address the accuracy challenges faced by one-stage dense object detectors due to class imbalances during training.
  • Developed a novel loss function called Focal Loss to down-weight well-classified examples.
  • Designed and trained a dense object detector called RetinaNet using Focal Loss.
  • Evaluated the effectiveness of Focal Loss against traditional cross entropy loss.
  • RetinaNet trained with Focal Loss matched the speed of existing one-stage detectors.
  • Surpassed the accuracy of all state-of-the-art two-stage detectors.
  • Demonstrated significant improvement due to effective handling of class imbalance.

Cite This Study

Lin et al. (2017) studied this question.

synapsesocial.com/papers/696015fd61a0d350516be86ehttps://doi.org/10.1109/iccv.2017.324
View Full Paper
Ask AI
Bookmark
Share

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

  1. 1Hybrid Classification-Regression Adaptive Loss for Dense Object Detection2024 · 1 citations
  2. 2Adaptive Binary Focal Loss: Enhancing Radiograph Image Classification With Balanced Specificity and Sensitivity2025 · 1 citations
  3. 3Dense Object Detection Based on De-Homogenized Queries2024
  4. 4Efficient and Accurate Object Detection with Asymmetric Progressive Semi-Decoupled Head and Harmonic Focal Loss2026
  5. 5A lightweight remote sensing small-object detection approach with scale-based dynamic loss and efficient multi-scale attention2026