MOCS: A Dataset and Benchmark for Detecting Moving Objects in Construction Sites
Description
MOCS dataset is a large-scale image dataset for detecting Moving Objects in Construction Sites. All the images in MOCS dataset are collected from real construction sites. There are 41,668 images in dataset and 13 categories of objects are annotated. For details of the dataset, readers are refer to our research article.
Update:
• Online evaluation competition is opened! (2021.9.14)
• Annotations are released! (2020.10.12)
• Images of test-set are released! (2020.10.10)
• Images of train-set are released! (2020.10.10)
• Images of val-set are released! (2020.05.10)
• Results of Mask Scoring RCNN is reported! (2020.05.01)
• Images and annotations will be released soon.
Categories in MOCS: Worker, Tower crane, Hanging hook, Vehicle crane, Roller, Bulldozer, Excavator, Truck, Loader, Pump truck, Concrete transport Mixer, Pile driver, Other vehicle.
Image Examples in MOCS
武汉“两山医院”视频检测结果：
Here is a video for detecting objects in construction sites. The original video is an aerial video of the construction of "Huoshenshan", "Leishenshan" hospital in Wuhan (which are used for COVID-19)
Online Evaluation:
We establish an online codalab competition, to give a way of evaluating your test set result.
You need to create an account on CodaLab. This will allow you to participate in this challenge.
You can get more information about this competition on the MOCS competition website.
Benchmark Results:
Object detection
Detectors | mAP | AP0.5 | AP75 | Aps | APm | APl | Inference speed |
YOLO-v3(Darknet53) | 39.045 | 65.590 | 41.658 | 9.595 | 26.824 | 51.031 | 27.03 |
SSD300(VGG16) | 36.076 | 59.282 | 38.026 | 4.279 | 19.299 | 51.198 | 21.28 |
Retinanet(ResNet50+FPN) | 50.014 | 72.844 | 53.604 | 14.766 | 38.046 | 62.398 | 6.80 |
FCOS(ResNet50+FPN) | 46.850 | 69.316 | 50.015 | 13.052 | 34.230 | 60.346 | 14.93 |
NAS-FPN(ResNet50) | 47.595 | 68.546 | 50.055 | 7.012 | 30.642 | 64.028 | 17.61 |
Tridentfast(ResNet50) | 50.686 | 73.077 | 54.714 | 15.335 | 37.784 | 63.632 | 4.05 |
Faster R-CNN(ResNet50+FPN) | 50.639 | 74.649 | 55.588 | 19.299 | 39.585 | 61.334 | 8.39 |
Faster R-CNN(ResNet101+FPN) | 50.672 | 73.982 | 55.311 | 18.296 | 38.837 | 62.483 | 6.79 |
Faster R-CNNResNeXt101+FPN) | 52.451 | 74.964 | 57.594 | 20.137 | 40.911 | 64.206 | 4.01 |
Instance Segmentation
Detectors | mAP | AP0.5 | AP75 | Aps | APm | APl | Inference speed |
SOLO(ResNet50+FPN) | 43.637 | 68.793 | 45.945 | 8.175 | 29.567 | 59.334 | 6.25 |
PointRend(ResNet50+FPN) | 47.906 | 73.149 | 52.287 | 14.979 | 34.001 | 61.516 | 8.86 |
BlendMask(ResNet50+FPN) | 44.876 | 70.804 | 47.309 | 12.618 | 32.383 | 58.608 | 8.41 |
Mask Scoring RCNN(ResNet50+FPN) | 42.858 | 66.847 | 45.984 | 13.501 | 31.965 | 56.198 | 9.70 |
Mask R-CNN(ResNet50+FPN) | 43.175 | 71.266 | 45.351 | 14.927 | 32.150 | 55.054 | 7.66 |
Mask R-CNN(ResNet101) | 45.352 | 72.743 | 47.985 | 15.381 | 33.643 | 57.593 | 6.22 |
Mask R-CNN(ResNeXt101) | 46.875 | 74.122 | 49.798 | 16.173 | 36.024 | 58.940 | 4.16 |
Download
Request form下载
Liscense
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
If you find the dataset useful, Please cite our paper.
If you want to use it for commercial purpose, please contact https://ott.tsinghua.edu.cn/
For communication or questions about dataset, please contact zhouli18@mails.tsinghua.edu.cn; sar20@mails.tsinghua.edu.cn
