AI Paper Atlas

Research map / Detection and segmentation / CNN features

Focal Loss for Dense Object Detection

Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, Piotr Dollar

ICCV 2017 · Detection and segmentation › CNN features · approach: Region-based and query-based detectors · 31,880 citations (OpenAlex)

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 changedWhereMeasuredEffect
use focal loss gamma=2 instead of alpha-balanced CE (ablation)RetinaNet ResNet-50-FPN · COCO minivalAP: 31.1 -> 34.0Improved
alpha-balance the cross entropy (alpha=.75 vs .50) (ablation)RetinaNet CE · COCO minivalAP: 30.2 -> 31.1Improved
raise focusing parameter gamma from 0 to 2 (ablation)Focal loss · COCO minivalAP: 31.1 -> 34.0Improved
raise gamma from 2.0 to 5.0 (ablation)Focal loss · COCO minivalAP: 34.0 -> 32.2Hurt
use focal loss instead of best OHEM (ablation)RetinaNet ResNet-101 · COCO minivalAP: 32.8 -> 36.0Improved
use 3 scales and 3 aspect ratios instead of one square anchor (ablation)RetinaNet FL · COCO minivalAP: 30.3 -> 34.0Improved
increase anchors from 9 to 12 (4 scales) (ablation)RetinaNet FL · COCO minivalAP: 34.0 -> 33.8No clear effect
add prior initialization pi=.01 to the last layer for CE (ablation)RetinaNet CE · COCOtraining stability: diverging -> 30.2 APImproved
increase ResNet depth from 50 to 101 at 600pxRetinaNet · COCO test-devAP: 34.3 -> 36.0Improved
increase image scale from 400 to 800 with ResNet-101RetinaNet · COCO test-devAP: 31.9 -> 37.8Improved
train with scale jitter and 1.5x longerRetinaNet-101-800 · COCO test-devAP: +1.3 APImproved
swap ResNet-101-FPN for ResNeXt-101-FPN backboneRetinaNet · COCO test-devAP: 39.1 -> 40.8Improved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy