Libra R-CNN Detection Model

[315386775]

Paper: Libra R-CNN: Towards Balanced Learning for Object Detection.
Link: https://arxiv.org/abs/1904.02701
Introduction: It integrates two novel components: balanced feature pyramid, and balanced L1 loss, respectively for reducing the imbalance at feature, and objective level. Benefitted from the overall balanced design, Libra R-CNN significantly improves the detection performance. Without bells and whistles, it achieves 2.5 points and 2.0 points higher Average Precision (AP) than FPN Faster R-CNN and RetinaNet respectively on MSCOCO.

[315386775]

Models of VOC dataset are evaluated with native resolutions with shorter side >= 600 but longer side <= 1000 without changing aspect ratios. This pr is balance loss version

detection module	map
faster_rcnn_resnet50_v1b_voc	78.3
libra_rcnn_balance	79.8
libra_rcnn_voc	80.9

[github-actions]

Job PR-1733-adc4358 is done.
Docs are uploaded to http://gluon-vision-staging.s3-website-us-west-2.amazonaws.com/PR-1733/adc4358/index.html

[bryanyzhu]

This PR looks good. May I ask what is the difference between this PR and PR #1727 ? It seems this PR doesn't have the corresponding model definitions. Should I close #1727 or merge both PRs? Thank you.

[315386775]

@bryanyzhu thanks for your reply. i got the bug with the code after i pull the code in the PR #1727 . i will close the #1727 and re-upload the balance fpn module.
This PR's pretrained model and log. i will upload to the google drive soon.