Inference: labels & classes

I’ve got a couple of questions on inferencing and generation of annotated images in the TAO detectnet_v2 sample i.e running:

!tao detectnet_v2 inference -e $SPECS_DIR/myconfig1.txt \
                            -o $USER_EXPERIMENT_DIR/myconfig1_infer_testing \
                            -i $DATA_DOWNLOAD_DIR/test_samples \
                            -k $KEY
  1. How can I create bounding box annotations for specific classes only. For example, I’ve trained a model for cats, dogs and apes. but only want annotated images with bounding boxes for cats. I’ve tried editing out various sections of the config file, without success

  2. What determines the if class name / confidence is shown on the bounding boxes ??

The sample detectnet_v2_inference_kitti_etlt_qat.txt seems to generate them, but not detectnet_v2_inference_kitti_tlt.txt, but I can’t see any relevant difference in the config file

thank you

Andrew

1)Please try to modify bbox_handler_config to draw the expected class.
2) Refer to DetectNet_v2 — TAO Toolkit 4.0 documentation

thanks @Morganh

I have already been through the documentation

  1. Do you mean my config should have target_classes
    in the inferencer_config section for cat, dog & ape but I hide some bboxes with a setting (such as a high threshold) for classwise_bbox_handler_config say for cats & dogs

  2. Which parameter in particular ? There is nothing in either the postprocessing_config or bbox_handler_config sections

Could you share your myconfig1.txt?

Hi @Morganh

I’ve attached my inferencing config & also a sample image/kitti output

thanks

Andrew

detectnet_v2_inference_kitti_tlt_steamranger_v7.txt (7.4 KB)

GX010098_frame_0000002310.txt (271 Bytes)

To visualize vad only, can you change bbox_handler_config part to below. Just delete other classes and last “default” setting.

bbox_handler_config{
  kitti_dump: true
  disable_overlay: false
  overlay_linewidth: 2


  classwise_bbox_handler_config{
    key:"vad"
    value: {
      confidence_model: "aggregate_cov"
      output_map: "vad"
      bbox_color{
        R: 26
        G: 147
        B: 52
      }
      clustering_config{
        clustering_algorithm: DBSCAN
        coverage_threshold: 0.005
        dbscan_eps: 0.3
        dbscan_min_samples: 0.05
        dbscan_confidence_threshold: 0.9
        minimum_bounding_box_height: 4
      }
    }
  }
}

Hi @Morganh

Thanks for the reply & sorry for the slow response.
I’ve since tried a Yolov4 model in TAO, rather than DetectNet & getting way better inference results- and the BBOXes are produced correctly in the inference output- so I think I’ll stick with this.

Andrew

Andrew