gengmufeng/MixtureOfDeepExperts

Implements a gated network for fusing features from different modalities for object detection

MixtureOfDeepExperts

This is implementation is loosely based on the paper "Choosing smartly". It incorporates multiple CNN detectors "Experts"and combines their output using a gated weighting network.

Sample outputs:

RGB channel	Depth Channel	Gating network

Dependencies :

• detectron2
• pytorch 1.7.0 + cuda 11.0
• InOutDoorPeople dataset
• download and extract the dataset from above link and modify the data path accordingly before executing.

Training :

To train CNN "Experts" Models over single modality use the below code, eg , RGB or Depth

usage: train_singleExpert.py [-h] [--modality_path MODALITY_PATH]
                             [--batch_size BATCH_SIZE] [--workers WORKERS]
                             [--iterations ITERATIONS] [--out_dir OUT_DIR]
optional arguments:
  -h, --help            show this help message and exit
  --modality_path MODALITY_PATH
                        path to RGB model
  --batch_size BATCH_SIZE
                        batch size for dataloader
  --workers WORKERS     no of workers for dataloader
  --iterations ITERATIONS
                        no. of iterations for training
  --out_dir OUT_DIR     output directory to save models

After training the CNN Expert models train the gating network using checkpoints from the above outputs

Training Gating network
usage: train_gatingNetwork.py [-h] [--model1 MODEL1] [--model2 MODEL2]
                              [--batch_size BATCH_SIZE]
                              [--no_of_workers NO_OF_WORKERS] [--data DATA]
                              [--epoch EPOCH] [--out_dir OUT_DIR]

optional arguments:
  -h, --help            show this help message and exit
  --model1 MODEL1       path to RGB model
  --model2 MODEL2       path to Depth model
  --batch_size BATCH_SIZE
                        batch size for dataloader
  --no_of_workers NO_OF_WORKERS
                        no of workers for dataloader
  --data DATA           path to InOutDoorData
  --epoch EPOCH         no. of epochs for training
  --out_dir OUT_DIR     output directory to save models

Evaluation :

Use the below commands to evaluate the trained models.

usage: eval_single.py [-h] [--modality_path MODALITY_PATH] [--data DATA]
                     [--batch_size BATCH_SIZE] [--workers WORKERS]
                     [--iterations ITERATIONS] [--out_dir OUT_DIR]
optional arguments:
 -h, --help            show this help message and exit
 --modality_path MODALITY_PATH
                       path to RGB model
 --data DATA           data directory to save models
 --batch_size BATCH_SIZE
                       batch size for dataloader
 --workers WORKERS     no of workers for dataloader
 --iterations ITERATIONS
                       no. of iterations for training
 --out_dir OUT_DIR     output directory to save models

usage: eval_gating.py [-h] [--model1 MODEL1] [--model2 MODEL2] [--gated GATED]
                      [--data DATA] [--out_dir OUT_DIR]
optional arguments:
  -h, --help         show this help message and exit
  --model1 MODEL1    path to RGB model
  --model2 MODEL2    path to Depth model
  --gated GATED      path to gated model
  --data DATA        path to InOutDoorData
  --out_dir OUT_DIR  output directory to save models