Visual contents, including images and videos, are dominant on the Internet today. The conventional search engine is mainly designed for textual documents, which must be extended to process and manage increasingly high volumes of visual data objects. In this paper, we present Mixer, an effective system to identify and analyze visual contents and to extract their features for data retrievals, aiming at addressing two critical issues: (1) efficiently and timely understanding visual contents, (2) retrieving them at high precision and recall rates without impairing the performance. In Mixer, the visual objects are categorized into different classes, each of which has representative visual features. Subsystems for model production and model execution are developed. Two retrieval layers are designed and implemented for images and videos, respectively. In this way, we are able to perform aggregation retrievals of the two types in efficient ways. The experiments with Baidu's production workloads and systems show that Mixer halves the model production time and raises the feature production throughput by 9.14x. Mixer also achieves the precision and recall of video retrievals at 95% and 97%, respectively. Mixer has been in its daily operations, which makes the search engine highly scalable for visual contents at a low cost. Having observed productivity improvement of upper-level applications in the search engine, we believe our system framework would generally benefit other data processing applications.
more »
« less
Mixer: efficiently understanding and retrieving visual content at web-scale
Visual contents, including images and videos, are dominant on the Internet today. The conventional search engine is mainly designed for textual documents, which must be extended to process and manage increasingly high volumes of visual data objects.In this paper, we present Mixer, an effective system to identify and analyze visual contents and to extract their features for data retrievals, aiming at addressing two critical issues: (1) efficiently and timely understanding visual contents, (2) retrieving them at high precision and recall rates without impairing the performance. In Mixer, the visual objects are categorized into different classes, each of which has representative visual features. Subsystems for model production and model execution are developed. Two retrieval layers are designed and implemented for images and videos, respectively.In this way, we are able to perform aggregation retrievals of the two types in efficient ways. The experiments with Baidu’s production workloads and systems show that Mixer halves the model production time and raises the feature production throughput by 9.14x.Mixer also achieves the precision and recall of video retrievals at 95% and 97%, respectively. Mixer has been in its daily operations, which makes the search engine highly scalable for visual contents at a low cost. Having observed productivity improvement of upper-level applications in the search engine, we believe our system framework would generally benefit other data processing applications,
more »
« less
- Award ID(s):
- 1718450
- NSF-PAR ID:
- 10294981
- Date Published:
- Journal Name:
- Proceedings of the VLDB Endowment
- Volume:
- 14
- Issue:
- 12
- ISSN:
- 2150-8097
- Page Range / eLocation ID:
- 2906-2917
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
With the advancement and dominant service of Internet videos, the content-based video deduplication system becomes an essential and dependent infrastructure for Internet video service. However, the explosively growing video data on the Internet challenges the system design and implementation for its scalability in several ways. (1) Although the quantization-based indexing techniques are effective for searching visual features at a large scale, the costly re-training over the complete dataset must be done periodically. (2) The high-dimensional vectors for visual features demand increasingly large SSD space, degrading I/O performance. (3) Videos crawled from the Internet are diverse, and visually similar videos are not necessarily the duplicates, increasing deduplication complexity. (4) Most videos are edited ones. The duplicate contents are more likely discovered as clips inside the videos, demanding processing techniques with close attention to details. To address above-mentioned issues, we propose Maze, a full-fledged video deduplication system. Maze has an ANNS layer that indexes and searches the high dimensional feature vectors. The architecture of the ANNS layer supports efficient reads and writes and eliminates the data migration caused by re-training. Maze adopts the CNN-based feature and the ORB feature as the visual features, which are optimized for the specific video deduplication task. The features are compact and fully reside in the memory. Acoustic features are also incorporated in Maze so that the visually similar videos but having different audio tracks are recognizable. A clip-based matching algorithm is developed to discover duplicate contents at a fine granularity. Maze has been deployed as a production system for two years. It has indexed 1.3 billion videos and is indexing ~800 thousand videos per day. For the ANNS layer, the average read latency is 4 seconds and the average write latency is at most 4.84 seconds. The re-training over the complete dataset is no longer required no matter how many new data sets are added, eliminating the costly data migration between nodes. Maze recognizes the duplicate live streaming videos with both the similar appearance and the similar audio at a recall of 98%. Most importantly, Maze is also cost-effective. For example, the compact feature design helps save 5800 SSDs and the computation resources devoted to running the whole system decrease to 250K standard cores per billion videos.more » « less
-
null (Ed.)Mobile apps are one of the most widely used types of software systems in existence today and more programmers and students learn how to develop them everyday. One of the most popular resources for learning mobile programming are videos hosted on social platforms such as YouTube. While useful, this type of resource has also its limitations, especially when developers are looking for user interface (UI) designs for mobile applications, since these are hard to search for and locate in videos. We propose UIScreens, a web-based analysis and search engine that analyzes the visual contents of mobile programming video tutorials, then identifies and extracts the UI screens displayed in the videos. Our tool offers features such as searching for UI screens in videos, displaying an overview of the UI screens identified in a video under each search result, and navigating to the part of a video where a particular UI screen is being displayed and discussed. In a user study, participants agreed that UIScreens is usable and useful to quickly skim through videos, while the UI screens it extracts can help developers further determine the relevance of videos to a search topic.more » « less
-
null (Ed.)Urban flooding is a major natural disaster that poses a serious threat to the urban environment. It is highly demanded that the flood extent can be mapped in near real-time for disaster rescue and relief missions, reconstruction efforts, and financial loss evaluation. Many efforts have been taken to identify the flooding zones with remote sensing data and image processing techniques. Unfortunately, the near real-time production of accurate flood maps over impacted urban areas has not been well investigated due to three major issues. (1) Satellite imagery with high spatial resolution over urban areas usually has nonhomogeneous background due to different types of objects such as buildings, moving vehicles, and road networks. As such, classical machine learning approaches hardly can model the spatial relationship between sample pixels in the flooding area. (2) Handcrafted features associated with the data are usually required as input for conventional flood mapping models, which may not be able to fully utilize the underlying patterns of a large number of available data. (3) High-resolution optical imagery often has varied pixel digital numbers (DNs) for the same ground objects as a result of highly inconsistent illumination conditions during a flood. Accordingly, traditional methods of flood mapping have major limitations in generalization based on testing data. To address the aforementioned issues in urban flood mapping, we developed a patch similarity convolutional neural network (PSNet) using satellite multispectral surface reflectance imagery before and after flooding with a spatial resolution of 3 meters. We used spectral reflectance instead of raw pixel DNs so that the influence of inconsistent illumination caused by varied weather conditions at the time of data collection can be greatly reduced. Such consistent spectral reflectance data also enhance the generalization capability of the proposed model. Experiments on the high resolution imagery before and after the urban flooding events (i.e., the 2017 Hurricane Harvey and the 2018 Hurricane Florence) showed that the developed PSNet can produce urban flood maps with consistently high precision, recall, F1 score, and overall accuracy compared with baseline classification models including support vector machine, decision tree, random forest, and AdaBoost, which were often poor in either precision or recall. The study paves the way to fuse bi-temporal remote sensing images for near real-time precision damage mapping associated with other types of natural hazards (e.g., wildfires and earthquakes).more » « less
-
Speech emotion recognition (SER) is a challenging task due to the limited availability of real-world labeled datasets. Since it is easier to find unlabeled data, the use of self-supervised learning (SSL) has become an attractive alternative. This study proposes new pre-text tasks for SSL to improve SER. While our target application is SER, the proposed pre-text tasks include audio-visual formulations, leveraging the relationship between acoustic and facial features. Our proposed approach introduces three new unimodal and multimodal pre-text tasks that are carefully designed to learn better representations for predicting emotional cues from speech. Task 1 predicts energy variations (high or low) from a speech sequence. Task 2 uses speech features to predict facial activation (high or low) based on facial landmark movements. Task 3 performs a multi-class emotion recognition task on emotional labels obtained from combinations of action units (AUs) detected across a video sequence. We pre-train a network with 60.92 hours of unlabeled data, fine-tuning the model for the downstream SER task. The results on the CREMA-D dataset show that the model pre-trained on the proposed domain-specific pre-text tasks significantly improves the precision (up to 5.1%), recall (up to 4.5%), and F1-scores (up to 4.9%) of our SER system.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- The DOI is not currently available.
