You are here

Improved Identification of the function of Novel and Partially Occluded Laboratory Equipment.

Description:

TECHNOLOGY AREA(S): Chem Bio_defense 

OBJECTIVE: Enable military operators to identify novel and partially occluded laboratory equipment 

DESCRIPTION: Military operators need to identify laboratory equipment they encounter in facilities (DTRA-17-BAA-RIF-0001)1. Currently, machine learning algorithms are being used to aid in object identification. However these algorithms are very dependent on the labelled training set used to train the models and can fail when real world conditions differ from the training set used (e.g. pristine pictures of laboratory equipment from a catalog vs pictures where objects are partially occluded). Additionally, traditional machine learning algorithm based approaches can fail if they encounter novel objects that might have the exact function but not appearance of an object used in the training set (e.g. two different brands of thermal cyclers). There is a need for improved methods to detect and classify objects which can overcome the challenges mentioned above. Recent advances in computer science may aid in overcoming the above challenges by taking into account the context in which the object appears in conjunction with other objects (e.g. co-occurrence, relative location, etc.)2,3 as well as background knowledge using approaches such as knowledge graphs4. For example a beaker with a stir bar and liquid inside it that is on top of an unidentified plate may have a higher probability of being a magnetic stir plate due to the relative location of the different objects. While developed methodologies will be tailored for use by DTRA for improved identification of laboratory equipment, it is expected they will have a broader potential customer base for any technology which requires object identification. Multidisciplinary teams composed of experts in areas such as machine learning, knowledge graphs, statistical based approaches, and the life sciences are preferred. Laboratory equipment shall be limited to equipment used in the life sciences area. Proposals should identify and explain the content of the data sources they propose to use. Teams should be self-sufficient and should not rely on DTRA to identify relevant relationships or provide data. 

PHASE I: The phase I deliverable is a report and preliminary proof of concept demonstration detailing the methods used for 1) improved identification of occluded life science laboratory equipment. The performer shall identify, collect, label, and identify relationships for 100 life sciences laboratory equipment. The performer shall develop metrics to measure performance. The report shall detail the (1) advantages and disadvantages/limitations of the proposed methods (2) benchmark data, and (3) preliminary proof of concept demonstration. 

PHASE II: The phase II deliverable is a final report and final proof of concept demonstration detailing the methods used for 1) improved identification of occluded life science laboratory equipment and 2) identification of the function of previously unseen (i.e.no labelled training data available) life sciences laboratory equipment. The performer shall identify, collect, label, and identify relationships for 1000 life sciences laboratory equipment. The performer shall continue to measure performance. The final report shall detail the (1) advantages and disadvantages/limitations of the proposed methods (2) benchmark data, and (3) demonstration. 

PHASE III: Finalize and commercialize software for use by customers (e.g. DTRA, industry). Although additional funding may be provided through DoD sources, the awardee should look to other public or private sector funding sources for assistance with transition and commercialization. 

REFERENCES: 

1: ​ https://www.fbo.gov/utils/view?id=51f3d824eac86a5d861dfce4415eb8b4#_Toc440615665

2:  ​Hong, Jongkwang, et al. "Discovering overlooked objects: Context-based boosting of object detection in indoor scenes." Pattern Recognition Letters 86 (2017): 56-61

3:  ​Guan, Linting, Yan Wu, and Junqiao Zhao. "SCAN: Semantic Context Aware Network for Accurate Small Object Detection." International Journal of Computational Intelligence Systems 11.1 (2018): 936-950

4:  ​Fang, Yuan, et al. "Object detection meets knowledge graphs." (2017)

KEYWORDS: Context Dependent Learning, Knowledge Graphs, Bayesian, Lab Equipment, Object Identification, Computer Vision 

US Flag An Official Website of the United States Government