This Repository is made to experiment with DINO object Detection pretrained model Link. Here, DINO object Detection pre-trained model is been fine-tuned on pedestrian dataset collected within the IIT Delhi Campus.

• As macos does not come with Nvidia CuDa , so the code is written in the Google colab environment.

• Get access to GPU .

• clone the repository and setup that on your environment.

• Install the dependencies and test the setup.

• this will require yaml version 0.31.0 , so install it through this command. !pip install yapf==0.31.0

• Download the pre-trained DINO-4scale model with the ResNet-50 (R50) from Here

• Download the Pedestrain dataset from Here

• setup the dataset in COCODIR format , As:-

  COCODIR/
    ├── train2017/
    ├── val2017/
    └── annotations/
    	├── instances_train2017.json
    	└── instances_val2017.json
  

• After setting up the directory and repository , Evaluate the model on the validation set and visualize the Results.

• Train / Fine-Tune the model with the training set and Re-evaluate the model and note the changes .

• One will need to make changes in the config file DINO/config/DINO_4scale.py to fine-tune it .

• To leverage the pre-trained models for model fine-tuning, i suggest add two more commands in a bash:

  --pretrain_model_path /path/to/a/pretrianed/model. specify a pre-trained model.
  --finetune_ignore label_enc.weight class_embed. ignore some inconsistent parameters.
  

After completing every steps , i made a small report which states the model working , analysis , challenges faced. LINK:- Report on DINO

• changed the weights of the pre-trained model.
• changed parameters like : num_classes , dn_labebook_size , name_id

1. DINO.ipynb :- This python file is the main file where i performed the task.
2. slice_annotaions.py :- With the help of this file one can slice the annotations in two part training and validation
3. loss_graphs.py :- This python file conisist of code which can help to generate the loss graph while fine-tuning the model.
4. Attention_map.py :- through this file one can generate the attention map of input images while going through training/fine-tuning.