{"id":57,"date":"2022-04-26T23:12:57","date_gmt":"2022-04-26T23:12:57","guid":{"rendered":"https:\/\/mscvprojects.ri.cmu.edu\/2022team9\/?page_id=57"},"modified":"2022-12-20T17:26:49","modified_gmt":"2022-12-20T17:26:49","slug":"results","status":"publish","type":"page","link":"https:\/\/mscvprojects.ri.cmu.edu\/2022team9\/results\/","title":{"rendered":"Results"},"content":{"rendered":"\n
\n

Approach 1: Detection Model trained on COCO, LVIS, NuScenes, ArgoVerse 2.0<\/strong><\/h2>\n\n\n\n

The results for the experiments run during the early stages of the project owing to approach 1<\/a> as explained in the previous section is discussed here. We ran the LVIS + NuScenes + COCO model on the 100 live cameras and collected data for a month and performed the following analysis.<\/p>\n\n\n\n

  1. We first got an understanding of the counts of the selected LVIS classes detected across the active cameras for a given time frame. The X-axis indicates the camera IDs and Y-axis indicates the counts and the colors indicate the LVIS class type.<\/li><\/ol>\n\n\n\n
    \"\"
    The plot shows the LVIS classes captured across all the active cameras [camera_id = range[1,104]] and their counts as detected per camera.<\/figcaption><\/figure>\n\n\n\n

    We see that a class like sign board is detected in every other camera and hence though a necessary condition, not a sufficient condition for presence of work zone scene. Hence we consider top – n cameras with highest counts for the cone (“brown” in the above plot) class and perform further analysis in them.<\/p>\n\n\n\n

    2. Now we further try to analyze the count distribution of each of these classes in the camera selected in the previous step. <\/p>\n\n\n\n

    \"\"
    LVIS class count across time on the camera that captured highest number of LVIS classes in that time period
    <\/figcaption><\/figure>\n\n\n\n

    As we can see from the plot, there is a high probability of finding a frame containing a work zone scene at and around the time stamp marked in red.<\/p>\n\n\n\n

    3. With the results from the analysis performed in the previous steps, we sample the corresponding frames from the camera footages and the results are as follows:<\/p>\n\n\n\n

    \"\"
    These are some of the work-zone scene frames we were able to sample using the steps mentioned above.<\/figcaption><\/figure>\n<\/div><\/div>\n\n\n\n\n\n
    \n

    Approach 2: Instance Segmentation Model trained on ILIM Dataset<\/strong><\/h2>\n\n\n\n

    Results corresponding to the approach 2<\/a> discussed in the previous page are shown here. A Mask-RCNN model was fine-tuned on the annotated dataset and analyzed for results. The model performed at an mAP of 0.324 with IOU threshold of 0.5. While the metrics of the trained model did establish a baseline performance for the task of construction zone detection, it was not satisfactory. <\/p>\n\n\n\n

    <\/p>\n\n\n\n

    Qualitative Results<\/h2>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n

    Takeaways<\/h2>\n\n\n\n