Before you start, please follow the instructions to prepare the dataset as described here.

If you have done this step in the refinement demo, skip this step.

Download the model checkpoint and coarse proposals here. The files in the zipped file have the following structure. ${V-A} denotes the unzipped file directory.

${V-A}
├── epoch_80.tar (The model checkpoint)
├── save_config.py (The experiment configuration file)
├── IDA-3D (predictions generated by IDA-3D)
   ├── xxxxxx.txt (IDA-3D predictions for image xxxxxx)
├── RTS-3D (predictions generated by RTS-3D)
   ├── xxxxxx.txt (RTS-3D predictions for image xxxxxx)      

Change directory to SNVC_DIR/tools and run

 python inference_agnostic.py --loadmodel ${V-A}/epoch_80.tar --output_dir OUTPUT_DIR --pred_dir ${V-A}/IDA-3D --btest BATCH_SIZE

Here OUTPUT_DIR is the directory storing output files and BATCH_SIZE denotes the batch size. If you have a GPU with 12GB memory, you can use BATCH_SIZE <= 10.

This command loads IDA-3D predictions, refines the 3D object predictions corresponding to Table 3 in the paper. You can also use the provided RTS-3D predictions.

If you want to visualize the predictions, set --debug.

To obtain quantitative evaluations, change directory to SNVC_DIR/tools/kitti-eval and run

 ./evaluate_object_3d_offline GT_DIR OUTPUT_DIR/all_parts

Here GT_DIR is the path to ground truth kitti labels.

Updating