Chainercv Versions Save

This release fixes a bug found in Faster R-CNN model (#266). No new feature is added.

This is the first release that supports Chainer v2. For Chainer v1 users, please use the latest version in version 0.4.*.

Changes

• Migrate to Chainer v2 (#243)

This is the last version with new features for Chainer v1. The v0.4* will be updated when there is a critical bug that needs to be fixed. For users using Chainer v2, please use versions v0.5.0 and higher.

Changed API

Important Axis-dependent values follow row-column order instead of column-row order.

• Convert convention used to describe shape of images from WH to HW (#234)
• Coordinates of bounding boxes are ordered as (y_min, x_min, y_max, x_max) instead of (x_min, y_min, x_max, y_max). (#246)
• Coordinates of keypoints are ordered as (y, x). (#235)
• Order of keyword arguments such as x_flip, y_flip for transforms/image are swapped. (#236)

Other changes

• Changed interface for eval_semantic_segmentation. They are now eval_semantic_segmentation_iou (#215, #217).
• Changed name of eval_detection_voc to eval_detection_voc_ap. (#232)

New features

• add_apply_prediction_to_iterator is added. (#229)
• Add generate_random_bbox (#219)

Enhancement

• DetectionVocEvaluator can now report AP for each class. (#242)
• Use DetectionVocEvaluator inside the train code for Faster R-CNN. (#213)
• Accept pretrained_model path for demos (#218)
• Add pretrained option to eval_voc07 (#231)

Bug fixes

• Fix the range of label in test_faster_rcnn_train_chain (#212)
• Fix readthedoc to compile properly by installing Cython (#214)
• Fix case of package names (#216)
• Fix requirements.txt for docs (#221)
• Add links.rst (#223)
• Fix SegNet example's evaluation protocol (#249, #255)
• Fix eval_voc_07 (#239)