LSTD

LSTD performs low-shot transfer detection within the Progressive Object Transfer Detection (POTD) framework by distilling object knowledge from a well-annotated source detector to improve a target detector's object detection performance using limited annotations.


Key Features:

  • Low-Shot Learning: Utilizes a small number of annotated examples to improve detection performance in target domains by transferring knowledge from a source detector.
  • Transfer Knowledge (TK) Module: Transfers learned features from a well-trained source model to a less-annotated target model to enhance detection capabilities.
  • Background Depression (BD) Module: Suppresses background responses to increase emphasis on relevant object regions within images.
  • Integration with SSD (Single Shot MultiBox Detector): Incorporates SSD as the region proposal network (RPN) to generate candidate object bounding boxes.
  • Weakly-Supervised Transfer Detection (WSTD) and Recurrent Object Labeling: Employs recurrent object labeling to handle weakly-labeled images and further refine the target detector using supervision from LSTD.
  • Progressive Object Transfer Detection (POTD) framework: Uses a staged approach that leverages prior knowledge and minimal annotations for progressive detector adaptation.

Scientific Applications:

  • Medical imaging: Adapts object detectors to new medical imaging targets with few annotated examples.
  • Wildlife monitoring: Detects novel species or objects in ecological imagery where annotations are scarce.
  • Rapid adaptation to new objects or conditions: Enables deployment of detectors in domains that require quick generalization from limited labeled data.

Methodology:

Two-stage transfer process: (1) Low-Shot Transfer Detection (LSTD) distills knowledge from a source detector with abundant annotations to enhance a target detector with few annotations; (2) Weakly-Supervised Transfer Detection (WSTD) applies a recurrent object labeling mechanism to weakly-labeled images, exploiting supervision from LSTD to further train the target detector.

Topics

Details

Tool Type:
command-line tool
Added:
11/14/2019
Last Updated:
12/22/2020

Operations

Publications

Chen H, Wang Y, Wang G, Bai X, Qiao Y. Progressive Object Transfer Detection. IEEE Transactions on Image Processing. 2020;29:986-1000. doi:10.1109/tip.2019.2938680. PMID:31502974.

PMID: 31502974
Funding: - National Basic Research Program of China: 2016YFC1400704 - National Natural Science Foundation of China: 61876176, U1613211, U1713208 - Shenzhen Basic Research Program: JCYJ20170818164704758