05:35
HADCV21-1200-Per-Vis-Shah - Part 7
-
Video
-
Table of contents
-
Video
- Description
- Transcript
- Discussion
About the talk
PeR-ViS: Person Retrieval in Video Surveillance using Semantic Description
About speaker
He is a Graduate Student studying Master of Computing and Innovation at The University of Adelaide. I hold a B.E. in Electronics and Communication Engineering from L. D. College of Engineering. In addition, he is also one of the founding members of Computer Vision Group – LDCE, which focuses on problems of Computer Vision & Deep Learning. In the past, he has worked with Mehul S. Raval and Hiren Galiyawala on Description Based Person Identification in Unconstrained Surveillance Video.
View the profileHello, my name is Wendy pager. I'm going to present of work, but he's our son retrieval in video surveillance using semantic description. The main motivation behind of what is in in the person related to fiction on persons are struggling the network, takes an image. As input, and it finds a similarity between the very image and the Sorcerer and scream and based on that. The correct person will be retrieved which has a major limitation. In fact, the effective as the network require super image, consider the case of a lost person where we
only know about. The person's opinions, that won't mention approach fails and smart description or however, it will be useful in these cases. Based on the simulations in or work. We started the problem of person retrieval in video surveillance with the semantic descriptions. The code name of Oliver is to use the computer vision techniques to fully automate, this task and retrieve, the person in challenging conditions by using minimal number of semantic description. Descriptors. I would have to watch Barbies is very simple and straightforward where it's divided into three stooges first.
We use mask rcnn, for instance, segmentation estimation and semantic description to narrow down the search PS4, determine the correct person. In the future. As we can see, Barbies uses the following sequence of height. So cute color. So close to type. So cute pattern, liquid color Outlook button. And the gender the main tree to calculate exit. Real height is to get the head and feet point from the instant segmentation boundaries and correspondingly execs checked
the Torso and legs veggies for Effectiveness. Beautiful viewers to look at your work and code for more in-depth information on how these have been done. Debbie doesn't we have used is too soft by research data set which Isabel only tested for this particular person table. By using the gasket between the densenet 161 model for different descriptors and the validation accuracy be at you for the data has been shown in the table. What performance mattress mattresses? We use the IOU intersection of a union where the ground truth,
or box, will be compared with the nitrous generator generator with an effective approach. And that's why we surpassed the previous a purchase order by 30, by a fair margin. Debbie doesn't has been divided into four categories, such as very easy, easy medium, and hard conditions such as ocular different occlusion or Buy computer. Make sure you presents, the average are you for the food? You want a sequence is based on the above categories. And we can notice that from William to her sickness has approached several student through the correct person. We also
showcase the localization accuracy. 441 test cases at a sequence. Has encouraged viewers to take a post to guess the number of sequences that are considered to be hard on William. We perform one of the ablation experiments by using different networks for 5, sequences for the different categories respectively as number as 604-1321, 23 and 28, which are considered to be very easy, easy, easy and medium and hard respectively. And Which all students are results of all different networks. Have you shown in the table? And as we can
see compared to other networks benzoate video and once the incident 161 performance, very better than the other other Netflix. Considering the easy and hard sequences. We showcased an Xbox account for each filter in the freezer in sequence. If the correct person Richard before the ending of the sequence, the approach will not follow the sequence order and provide the right person with coordinates for the police presence in the surveillance footage, supposed to find the correct person and provides a
false negative at the end as shown in the image in the images. Thereby, we can rule, I mentioning the major benefit of our work that the selection of this order and sequence is easily defensible to order or more dense or small Stars, such as bag gap, on her diaper truck and where do heavy drop his present when it is. One of the important use of instant segmentation in the this problem as it allows precise height estimation and Patch action photos of and leg. We also believe that there is a more work
that needs to be done and the possible future involves the use of poses to mention the use of trekking in this the particular what and other descriptors to improve the performance for more real-world scenarios. We thank you all the time. Every thanks audience for listening to over. And we also feel that we also let him know to feel free to visit overall project, which is shown in the slight. Thank you again.
Buy this talk
Ticket
Interested in topic “IT & Technology”?
You might be interested in videos from this event
Similar talks
Buy this video
Conference Cast
With ConferenceCast.tv, you get access to our library of the world's best conference talks.
