The final results of the "Underwater Object Detection Algorithm Competition" are out! 60 days of competition, many highlights
There will also be live competitions and award ceremonies in the future, so stay tuned.
Text | Xiao Man
Before May Day, the results of the final defense of the first phase of the "Underwater Object Detection Algorithm Competition" of the 2020 National Underwater Robot (Zhanjiang) Competition were announced.
This "Underwater Target Detection Algorithm Competition" is divided into two events: "Underwater Optical Image Target Detection" and "Underwater Acoustic Image Target Detection", each with 1 first prize, 5 second prizes, and 8 third prizes. The results of the finals have been officially confirmed, and the specific award-winning situation is shown in the figure:
Note: The picture shows the list of winners of the final of the "Underwater Optical Image Target Detection" competition
Note: The picture shows the list of winners of the final of the "Underwater Acoustic Image Target Detection" competition
After two months of fierce competition, the MaybeTomorrow team won the first prize in "Underwater Optical Image Target Detection" and the passi0n team won the first prize in "Underwater Acoustic Image Target Detection".
New algorithm competition with many highlights
On February 28, 2020, Pengcheng Laboratory launched the "Underwater Target Detection Algorithm Competition".
This algorithm competition is the first phase of the "National Underwater Robot (Zhanjiang) Competition" jointly organized by the National Natural Science Foundation of China, Zhanjiang Municipal People's Government and Pengcheng Laboratory. The competition includes three activities: online competition, on-site competition and award ceremony. The "Underwater Object Detection Algorithm Competition" is one of the online competitions.
This competition is a brand-new algorithm competition. From registration to the final defense, the entire process is held online, which is very exciting. From the beginning of registration, the contestants are very enthusiastic; from the perspective of the competition system, the subjective and objective review mechanism highlights the rigor; during the competition, online submission, online automatic scoring, online dynamic rankings, online lectures, online recurrence speed test, Pengcheng Cloud Brain and Octopus system are added to the competition.
Diversity of players
A total of 1,811 teams registered for the dual events of this online competition, attracting students from 268 universities and colleges around the world and big data and algorithm talents from 382 companies, with an impact reaching 12 overseas countries.
During the competition, the algorithm contestants were very enthusiastic, and the algorithm competition works were submitted nearly 10,000 times in total. The competition in the dual competition rankings was very fierce.
Rigorous competition system
The preliminary round review adopts a combination of subjective and objective review mechanisms. Taking the optics competition as an example, in the objective review, the competition requires that the contestant's algorithm model must perform well in both "accuracy" and "speed"; in the subjective review, the reviewers designated by the two competitions will review the algorithm model materials and related codes submitted by the contestants online and make validity judgments to effectively identify whether the contestants have committed violations such as plagiarism and manual labeling.
At the same time, both events adopt "one-on-one" (one group of judges for one team) online video defense, and each team has a guiding teacher to provide guidance for the defense; more than 10 judges for each event prepare targeted questions in advance, and ask and analyze the defense team's defense materials and details of the preliminary round assessment, making the defense process very fair, efficient and smooth.
Expert online lectures
The organizers opened a learning column highly relevant to the competition topics for the contestants and organized two online lectures, in which three expert teachers explained the competition strategies in the field of underwater image target detection to the contestants and answered questions online, providing the contestants with detailed learning materials and guiding them to continuously optimize the algorithm models to improve their competition results.
Pengcheng Cloud Brain Support
In the evaluation of this algorithm competition, Pengcheng Cloud Brain and Octopus system were used as support for the reproduction and evaluation phase.
The algorithm speed test was completed on the resource allocation and scheduling management cluster of Pengcheng Cloud Brain and Octopus system. The system was fully debugged before the competition to ensure the smooth implementation of the algorithm speed test.
Accurate computing power and schedule control: Algorithm speed evaluation requires players to have a consistent computing environment. To this end, the organizer conducted a benchmark performance test on the computing resources allocated to each team by the Cloud Brain system, and marked abnormal test nodes to prevent abnormal resource access. At the same time, the fairness and justice of the competition was ensured by accurately controlling the process and time points of players' access to prepare for the algorithm speed evaluation.
It is worth mentioning that in order to help the contestants get familiar with the Pengcheng Cloud Brain system, the organizers provided a lot of introductory guidance materials about the Cloud Brain system, including container startup, environment installation, data mounting, exception handling and usage precautions.
60 days of competition to crown the champion
This "Underwater Target Detection Algorithm Competition" mainly focuses on breaking through technical bottlenecks in the field of underwater target detection algorithms. "Optical image target detection" needs to detect the locations of different seafood, and "Acoustic image target detection" mainly detects characteristic targets composed of special topography and man-made objects on the seabed.
After 60 days of fierce competition, multiple rounds of review and evaluation by expert judges, and final confirmation by the algorithm competition technical committee, the results of the two events, "Optical Image Target Detection" and "Acoustic Image Target Detection", have been announced.
Among them, the MaybeTomorrow team won the first prize in "Underwater Optical Image Target Detection" and the passi0n team won the first prize in "Underwater Acoustic Image Target Detection".
In the optical imaging competition, the MaybeTomorrow team took advantage of the scarcity of real underwater data and the differences in the distribution of training and test data. They used Cascade RCNN as the main model, combined it with cutting-edge deformable convolution and non-local modules, and innovatively used the idea of Auto-ML to automatically search for data augmentation solutions suitable for underwater target detection problems, thereby significantly improving the performance of the detection algorithm. This fully demonstrated their proficiency in basic detection models and their ability to rationally use cutting-edge machine learning algorithms.
In the acoustic image competition, the passi0n team analyzed the size distribution and aspect ratio distribution of the targets in the dataset in a characteristic way, successfully set the anchor box hyperparameters, and introduced the ResNet50 network with variable convolution to better adapt to the diversity of target shape distribution. In addition, its high-resolution backbone network Hrnet18 can learn rich high-resolution representations and avoid information loss of short-edge features; the semi-supervised method uses pseudo labels to fine-tune the model, which improves the robustness of the model. Innovations have been made in both data analysis and network fusion and adjustment.
So far, the first stage (online competition) of the National Underwater Robot (Zhanjiang) Competition has come to an end. At the same time, the on-site competition of the competition has also kicked off. The on-site competition (finals) will be held in Zhanjiang on May 18-19. The award ceremony for the two stages of the online and on-site competitions will be held on the 21st. Stay tuned.
Previous recommendations
*Recruitment:
Leiphone.com is urgently recruiting senior editors for the hot topics group. Interested parties please send your resume to liufangping@leiphone.com
京公网安备 11010802033920号