Skip to content

Latest commit

 

History

History
81 lines (59 loc) · 3.42 KB

README.md

File metadata and controls

81 lines (59 loc) · 3.42 KB

C6 Team 5 - Week 4

Directory structure

The Week 4 directory should have the following structure:

data
│
└───aic19-track1-mtmc-train
│
models
│
└───feature_extraction
└───object_detection

In the data folder, the aic19-track1-mtmc-train dataset should be stored. The models folder should contain the models for feature extraction and object detection.

Measuring velocity

To execute the speed estimation, the velocity.py file should be executed:

python velocity.py

Make sure you have the video files and the pretrained YOLO model in the same directory as velocity.py

Multi-camera tracking

Training the Triplet Loss model for feature extraction

First, generate the triplets for the training of the Triplet Loss model. To do this, the build_triplets.py file should be executed:

python build_triplets.py

That will generate 1200 triplets for each sequence in the dataset. The triplets will be stored in the data folder at /triplets and they will be combined; it will combine S03 and S04 at /S0304 to train a model for S01, S01 and S04 at /S0104 to train a model for S03, and S01 and S03 at /S0103 to train a model for S04.

After generating the triplets, the siamese_network.py file should be executed to train the Triplet Loss model, here one run example (do not add parameters if you want to use the default ones):

python script_name.py --mode train --combination 0 --lr 0.001 --weight_decay 0.0001 --batch_size 16 --epochs 20

Where:

  • --mode train: Sets the script to training mode.
  • --combination: Selects the first sequence combination for training. Index for selecting sequence. 0: S0103 data for S04, 1: S0104 for S03, 2: S0304 for S01.
  • --data_path: Specifies the base directory for the dataset.
  • --models_path: Specifies the directory to save trained models.
  • --lr: Sets the learning rate.
  • --weight_decay: Sets the weight decay for regularization.
  • --batch_size: Sets the batch size for training.
  • --epochs: Sets the number of epochs for training.
  • --margin: Sets the margin for the Triplet Loss function.

For testing the model, the siamese_network.py file should be executed with the following arguments:

python script_name.py --mode test --combination 0 --batch_size 16

Where you can compute the mean distance between the anchor and the positive samples and the mean distance between the anchor and the negative samples for a test set.

Performing multi-camera tracking

To execute the multi-camera tracking, the main.py file should be executed with the following arguments:

python main.py --config path_to_config

The path_to_config should be the path to the configuration file that contains the parameters for the multi-camera tracking. The configuration file should be a .yaml file with the following structure:

sequence: path_to_sequence
od_model: path_to_object_detection_model
feature_model: path_to_feature_extraction_model

The sequence parameter should be the path to the sequence that will be used for the multi-camera tracking. The od_model parameter should be the path to the object detection model that will be used for the multi-camera tracking. The feature_model parameter should be the path to the feature extraction model that will be used for the multi-camera tracking.

The results of the tracking will be stored in each of the camera folders in the sequence folder with the name 'camera_name_eval.txt'.