We use mAP at various thresholds on the front camera image for evaluation. To our knowledge, there is only one significant work in weakly-supervised 3D object detection VS3D which uses an in-network proposal generation and cross-modal transfer learning between camera images and point clouds. There are a couple of methods that use selective search to generate proposals like PCL, and OICR[5]. Although our method is bottlenecked due to offline proposal generation, our model is simple and is backed by solid geometric computer vision concepts. We are not evaluating 3D mAP because most bounding boxes are inaccurate due to occlusion length. This is not a big issue in terms of self-driving cars because the min_y of the bounding box i.e., the nearest distance of the obstacle in the direction of length, is accurate. max_y is highly affected by occlusion.

Comparison with Baselines

                                                           ***Mean average precision (AP) of Car on the KITTI validation set of variation Weakly supervised Methods taken from VS3D***

NB: The results other than ours are evaluated on KITTI validation data, which takes time to get evaluated as we have to submit the code to KITTI's official website. Our mAP is on a validation set split from the training set. We are confident these results would hold good for the KITTI validation set as our model is weakly supervised i.e. proposal generation is irrespective of the network. There is not much chance of overfitting because the KITTI train data itself is very diverse.

Comparison of losses

The difference in AP decreases as IoU increases, but overall Focal loss performs much better

Precision-Recall Curves

Precision-Recall curve of Car class with different IoU (legend)

Train loss curve

For training with focal loss

For training with BCE loss

Camera Results