Skip to content

week5.md

Latest commit

 

History

History
49 lines (39 loc) · 2.66 KB

week5.md

File metadata and controls

49 lines (39 loc) · 2.66 KB

Week 5 - Object Detection

Abstract

The goal of this week is to detect objects (e.g. car, truck, pedestrian, cyclist, traffic sign) in the images, specifying their position with a bounding box, and recognize their category. This is known as object detection. For this task, we used the maskrcnn-benchmark framework, open sourced by Facebook Research. The different models, fine-tuned from ImageNet, have been trained on a set of provided datasets: TT100K, Udacity, and KITTI. In addition, the performance of the models has been boosted by tuning the hyperparameters and performing data augmentation. The performance of the trained models has been measured using Average Precision (AP) and Average Recall (AR).

Completed tasks

  • (a) Train an existing object detection network
  • (b) Read 2 papers about object detection networks
  • (c) Train the network(s) on a different dataset
  • (d) Boost the performance of the network(s)
  • (e) Report showing the achieved results
    • README
    • Slides
    • Report

Implementation

  • Run the provided code and train the following networks on COCO:
    • Mask R-CNN
    • RetinaNet
  • Train each network with the following datasets:
    • TT100K
    • Udacity
    • KITTI
  • Boost the performance of the networks:
    • Hyperparameters tuning
    • Data augmentation

Results

Training results:

network dataset AP AP@50 AP@75 AR@1 AR@10 AR@100 FPS
Mask R-CNN coco_2014_minival 0.370 0.586 0.403 0.306 0.477 0.500 10.20
Mask R-CNN tt100k_valid 0.587 0.837 0.670 0.595 0.610 0.610 13.87
Mask R-CNN tt100k_test 0.764 0.962 0.920 0.791 0.812 0.812 14.57
Mask R-CNN udacity valid 0.209 0.408 0.188 0.169 0.306 0.308 14.72
Mask R-CNN udacity_test 0.230 0.443 0.214 0.179 0.323 0.325 14.67
Mask R-CNN kitti_valid 0.627 0.906 0.718 0.331 0.672 0.679 15.73
RetinaNet coco_2014_minival 0.367 0.559 0.393 0.320 0.503 0.528 9.00
RetinaNet tt100k_valid 0.524 0.777 0.583 0.600 0.625 0.625 9.73
RetinaNet tt100k_test 0.767 0.965 0.923 0.798 0.820 0.820 14.58
RetinaNet udacity valid 0.227 0.442 0.201 0.187 0.342 0.348 14.10
RetinaNet udacity_test 0.238 0.461 0.224 0.190 0.353 0.363 14.10
RetinaNet kitti_valid 0.565 0.851 0.630 0.298 0.615 0.628 15.53