[关闭]
@zhenni94 2015-12-12T20:17:20.000000Z 字数 3117 阅读 988

Simultaneous Detection and Segmentation

Paper: http://www.eecs.berkeley.edu/Research/Projects/CS/vision/papers/BharathECCV2014.pdf

Task explanation

Detect all instances of a category in an image and, for each instance, mark the pixels that belongs to it.

Algorithm Steps

  1. Proposal generation
  2. Feature extraction
  3. Region classification
  4. Region refinement

Evaluation Metrics

Output: a set of object hypotheses, with a predicted segmentation and a score.
Correct hypothesis: its segmentation overlaps with segmentation of a ground truth instance by more than 50%.
Penalize duplicates (as traditional bounding box task)

Metric:
PR: precision recall curve
AP: average precision (area under the curve)

Notation:
: computed in SDS way. (Messures accuracy of segmentation and need to get each instance separately and completely)
: computed in classical bounding box task

: the volumn under surface varying the threshold(accurate segmentation)

Proposal generation

Care about segments, not just boxes.
Use MCG: outperform on the object level Jaccard index metric.(meassures the average best overlap achieved by a condidate for ground truth object)
improved ~ 0.7% if use MCG instead of Selective Search

Feature extraction

(=> Use b-box of MCG region instead of Selective Search b-box; Use penultimate fully connected layer)

feature extractors

SDS_net

Region classification

Region refinement

添加新批注
在作者公开此批注前,只有你和作者可见。
回复批注