Optimizing labelling on radar-based grid maps using active learning

Automatic Label Creation Framework for FMCW Radar Images Using Camera Data

“…The results obtained in [11], as well as the results from [2] (if we include the tracking system to study multiple frames), achieved a higher accuracy than our proposed framework. However, [11] is able to achieve these results through an iterative approach where human interaction is still required to manually label part of the dataset in each iteration. At the same time, the technique was not tested with multi-class data, in contrast with our technique which is able to detect multiple targets in the same frame.…”

“…On the other hand, when using the proposed autolabeling framework, the required time for the same task was reduced to 6 minutes (us- Technique initial mAP Average speed Our approach 82.056% 3 frames/sec K. Patel et al [2] 65.30% 2 frames/sec F. Nobis et al [1] 57.50% -T.Y. Lim et al [16] 73.5% 40 frames/sec T. Winterling et al [11] 94.93% - ing a GPU platform). Therefore, the time reduction achieve in this dataset was 96.76% respect to the manual labeling process.…”

“…Focus on the autolabeling process, as in our paper, T. Winterling et al [11] proposed a technique based on a CNN to automatically generate labels for occupancy grid maps generated from radar data. In [11], the initial data is preprocessed to obtain an occupancy grid where each cell contains the information of the probability of an object being at that position. These maps are fed into a CNN where the author had to manually label a set of data in each iteration to improve the accuracy achieved.…”

“…These techniques aim to reduce the human errors in the labeling process as well as the time required for this task. Some examples of these techniques are proposed in [11] [12] [13] [14] [15], which are deeply analyzed in Section 2. Some of these techniques are based on a Sensor Fusion approach such as [1].…”

Automated Ground Truth Estimation of Vulnerable Road Users in Automotive Radar Data Using GNSS

“…One solution to this problem is to find models that require less data to converge, i.e., using classification models with small amount of training coefficients as in [1]. Moreover, techniques such as active learning were found to be able to reduce the required amount of data for radar-based classification tasks [2].…”