近年來，由於用於識別和追蹤道路上物體的深度學習技術獲得巨大進步，自動駕駛汽車也得到了巨大的發展。為了將深度學習技術運用 於系統上，在此之前，通常需要大量的已標註的影片來訓練神經元網絡 模型。然而，標註影片的過程往往非常耗時且乏味，目前主要依賴於人 工。自動化標註影片這個過程實際上是雞與蛋的問題:我們需要一個完 美的物體偵測和追蹤工具來標註影片，以便訓練完美的物體偵測和追 蹤演算法。一個可行的替代方案是使用尚未完美的工具標註影片，然後 手動更正結果。在此論文中，我們介紹了這種用於自動駕駛的半自動視 頻標記工具。我們的工具基於開源影片標註系統 VATIC。首先使用多對 象偵測器和追蹤器來註釋視頻。識別對象標記中的可能錯誤，然後將其 呈現給人類以產生正確的標註。在標註測試影片實驗中，結果顯示我們 的工具可以更快地完成影片標註任務，同時保持相同的標註質量。這套 半自動的影片標註工具是從開源工具 VATIC 修改而來的，可以從 Github 獲得。

In recent years, the development of autonomous cars has gained great momentum due to the vast advances in deep learning technique for recognizing and tracking objects on the roads. To apply the deep learning technique, a large set of properly annotated videos are normally needed to train the neuron network model. However, the process of annotating videos is very time-consuming and tedious, and currently it relies mainly on human. Automating this process is actually a chicken-and-egg problem: we need a perfect object detection and tracking tool to annotate the videos so as to train a perfect object detection and tracking algorithm. A viable alternative is to annotate the videos using a less perfect tool and then correct the results manually. In this paper, we introduce such a semi-automatic video labeling tool for autonomous driving. Our tool is based on the open-source video annotation system VATIC. A multi-object detector and tracker is first used to annotate the video. Possible errors in the labeling of the objects are identified and then presented to human annotators to produce correct annotations. Experiments on labeling test videos show that our tool can complete the annotation task faster, while maintaining the same quality as a human annotator. The proposed tool is modified from the open-source tool VATIC and is available from Github.

1 Introduction 1
2 Related Work 5
2.1 Tools without Trained Model ...................... 5
2.2 Tools with Trained Model ........................ 6
3 Methodology 8
3.1 Label Information in VATIC....................... 9
3.2 Handling the Label Switch Case..................... 10
3.3 Handling the False Positive Case .................... 12
3.4 Handling the Disappear Case ...................... 13
3.5 Handling the Displacement Case..................... 15
4 Evaluation 18
4.1 Performance Measuring.......................... 18
4.2 Workload Saving ............................. 20
4.3 Effects of Revision Ordering ....................... 23
4.4 Effects of Object Detection........................ 23
5 Conclusion 26

[1] VATIC autolabel features, 2018. https://github.com/billy0059/VATIC_ AutoLabelFeaturs.
[2] A. Ambardekar, M. Nicolescu, and S. Dascalu. Ground truth verification tool (gtvt) for video surveillance systems. pages 354–359, Feb 2009.
[3] Keni Bernardin and Rainer Stiefelhagen. Evaluating multiple object tracking performance: The clear mot metrics. EURASIP Journal on Image and Video Processing, 2008(1):246–309, May 2008.
[4] F. Comaschi, S. Stuijk, T. Basten, and H. Corporaal. A tool for fast ground truth generation for object detection and tracking from video. pages 368–372, Oct 2014.
[5] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013.
[6] I. Kavasidis, S. Palazzo, R. Di Salvo, D. Giordano, and C. Spampinato. A semi- automatic tool for detection and tracking ground truth generation in videos. pages 6:1–6:5, 2012.
[7] J. S. Lopez-Villa, H. D. Insuasti-Ceballos, S. Molina-Giraldo, A. Alvarez-Meza, and G. Castellanos-Dominguez. A novel tool for ground truth data genera- tion for video-based object classification. In 2015 20th Symposium on Signal Processing, Images and Computer Vision (STSIVA), pages 1–6, Sept 2015.
[8] David Mihalcik and David Doermann. The design and implementation of viper. 2003.
27
[9] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. CoRR, abs/1506.01497, 2015.
[10] Paolo Napoletano Raimondo Schettini Simone Bianco, Gianluigi Ciocca. An interactive tool for manual, semi-automatic and automatic video annotation. pages 88–99, 2015.
[11] Carl Vondrick, Donald Patterson, and Deva Ramanan. Efficiently scaling up crowdsourced video annotation. International Journal of Computer Vision, pages 1–21, 2012. 10.1007/s11263-012-0564-1.
[12] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. 2015.
[13] C. M. Yang, Y. Choo, and S. Park. Semi-automatic image and video annotation system for generating ground truth information. pages 821–824, Jan 2018.
[14] Jenny Yuen, B. Russell, Ce Liu, and A. Torralba. Labelme video: Building a video database with human annotations. pages 1451–1458, Sept 2009.