SSF3D: Strict Semi-Supervised 3D Object Detection with Switching Filter

TL;DR

SSF3D tackles semi-supervised 3D object detection on LIDAR point clouds by introducing a strict thresholding mechanism (GMM3 Picker) and a dynamic entropy-based filter switching to minimize incorrect pseudo-labels while maintaining sufficient positives. The approach employs a two-stage online-offline training framework with Dense Loss and a point-removal/addition strategy to refine pseudo-label quality and learning signal. Evaluations on KITTI with 1% and 2% labeled data show meaningful gains over prior SS3DOD methods, including substantial improvements for the Cyclist class and consistent gains for Car, demonstrating strong few-shot capabilities. The findings suggest that combining strict, data-driven filtering with information-theoretic label quality assessments can effectively curb false negatives and label noise in semi-supervised 3D detection, with practical implications for reducing annotation costs.

Abstract

SSF3D modified the semi-supervised 3D object detection (SS3DOD) framework, which designed specifically for point cloud data. Leveraging the characteristics of non-coincidence and weak correlation of target objects in point cloud, we adopt a strategy of retaining only the truth-determining pseudo labels and trimming the other fuzzy labels with points, instead of pursuing a balance between the quantity and quality of pseudo labels. Besides, we notice that changing the filter will make the model meet different distributed targets, which is beneficial to break the training bottleneck. Two mechanism are introduced to achieve above ideas: strict threshold and filter switching. The experiments are conducted to analyze the effectiveness of above approaches and their impact on the overall performance of the system. Evaluating on the KITTI dataset, SSF3D exhibits superior performance compared to the current state-of-the-art methods. The code will be released here.

Figures (5)

• Figure 1: The proposed SSF3D result compare with state-of-the-art SS3DOD algorithmswang3DIoUMatchLeveragingIoU2021liuHierarchicalSupervisionShuffle2023a using only 2% labeled KITTIgeigerVisionMeetsRobotics2013 data, and full-dataset supervised result is given in gray bars.
• Figure 2: Threshold picking of different methods. The confidence thresholds selected by different SS3DOD methods in recent yearswang3DIoUMatchLeveragingIoU2021parkDetMatchTwoTeachers2022liDDS3DDensePseudoLabels2023liuHierarchicalSupervisionShuffle2023a are listed here, noted that those value are for car class, and means the lowest predictive score of the boxes that will be retained as labels. It can be seen that except our method, all other architecture adopt a moderate threshold to make a tradeoff between label quantity and quality, meanwhile introduce some incorrect labels. The * methods uses a point remove strategy.
• Figure 3: Ambiguous pseudo labels retained by confidence filters. In previous works, some ambiguous targets that have same points distribution with sundries are remained.
• Figure 4: Custom filter in 2 stages. Pipeline of the label filter in SSF3D. in stage-1 training, IOU FilterliuHierarchicalSupervisionShuffle2023a as Fig. \ref{['Filter']} (a) is used. In stage-2 we offline count all prediction boxes' 3-dimension-entropy as Fig. \ref{['Filter']} (b) and use the scatterplot of entropy to find a correct range to filter the generated labels.
• Figure 5: Scatterplot of all prediction boxes. Each points in the scatterplot represents a 1-stage-trained teacher model prediction box. The darker the red color, the higher the $IOU_{gt}$, and the dark blue indicates a classification error. (a) The scatterplot of entropy score in target's x,y,z dimension. (b) same scatterplot as Fig.(a) but with $IOU_{gt}$ color. (c) The scatterplot of confidence, roi and x-axis entropy score with $IOU_{gt}$ color.