Added readme

readme.md

@@ -0,0 +1,15 @@
+# Orthographic Feature Transform for Monocular 3D Object Detection
+
+This is a PyTorch implementation of the OFTNet network from the paper [Orthographic Feature Transform for Monocular 3D Object Detection](https://arxiv.org/abs/1811.08188). The code currently supports training the network from scratch on the KITTI dataset - intermediate results can be visualised using Tensorboard. The current version of the code is intended primarily as a reference, and for now does not support decoding the network outputs into bounding boxes via non-maximum suppression. This will be added in a future update. Note also that there are some slight implementation differences from the original code used in the paper. Please see `train.py` for details of training options.
+
+
+## Citation
+If you find this work useful please cite the paper using the citation below.
+```
+@article{roddick2018orthographic,  
+  title={Orthographic feature transform for monocular 3d object detection},  
+  author={Roddick, Thomas and Kendall, Alex and Cipolla, Roberto},  
+  journal={British Machine Vision Conference},  
+  year={2019}  
+}
+```

train.py

@@ -150,12 +150,7 @@ def compute_loss(pred_encoded, gt_encoded, loss_weights=[1., 1., 1., 1.]):
     score_weight, pos_weight, dim_weight, ang_weight = loss_weights
 
     # Compute losses
-    # score_loss = hard_neg_mining_loss(score, labels)
-    # score_loss = huber_loss(score, labels.float())
     score_loss = oft.model.loss.balanced_cross_entropy_loss(score, labels)
-    # # score_loss = oft.model.loss.heatmap_loss(score, labels, pos_weight=100)
-    # score_loss = oft.model.loss.uncertainty_loss(score, sqr_dists)
-    # # score_loss = oft.model.loss.log_ap_loss(score, sqr_dists)
 
     pos_loss = huber_loss(pos_offsets, gt_pos_offsets, labels.unsqueeze(2))
     dim_loss = huber_loss(dim_offsets, gt_dim_offsets, labels.unsqueeze(2))