Objective : To improve the base line code and achieve better results than current SOTA(please see this paper for the accuracy that we have to achieve "Rethinking of Pedestrian Attribute Recognition: Realistic Datasets and A Strong Baseline" - https://arxiv.org/pdf/2005.11909.pdf )
Base line code - https://github.com/valencebond/Strong_Baseline_of_Pedestrian_Attribute_Recognition
Base line codes paper - "Rethinking of Pedestrian Attribute Recognition: Realistic Datasets and A Strong Baseline" https://arxiv.org/pdf/2005.11909.pdf
All Datasets link - https://github.com/wangxiao5791509/Pedestrian-Attribute-Recognition-Paper-List
Literature survey - All papers related to it - https://github.com/wangxiao5791509/Pedestrian-Attribute-Recognition-Paper-List
Another base line code (but its not working with Rapv2 dataset) - https://github.com/dangweili/pedestrian-attribute-recognition-pytorch
Current Accuracies: (Taken from https://arxiv.org/pdf/2005.11909.pdf) 
Methods or different models available : Example : 
https://arxiv.org/pdf/2005.11576.pdf
Datasets that we are going to focus on : PA100k RAPv2 PETA RAPv1 The school of AI dataset - will send it later Market 1501 (optional)
Project Schedule : (Even if planned schedule didnt work its not a problem we can try to achieve it else we can reschedule)
Since we have a base line code and datasets already we need to improve the model alone thats enough . Each one can take a dataset and work on improving a model for that if any one gets a improvement in any one dataset we can proceed with that model .
Current baseline is on Resnet50 so we can think of using inception ,densenet or any other suitable network .
Literature survey 4 days
I think 12 days is enought for find a new model . so with in August 2 lets try to find a model that better than current model
So in august month 1- 15 we can write our paper .
July 17 - July 20 - Literature survey
July 21 - Aug 2 - Model preparation
Aug 2 - Aug 20 - Training part
Aug 20 - Aug 30 - paper writing
Accuracy to be achieved :
Name - Precision
Pa100k - 89.41
Rapv2- 81.99
PETA - 86.99
Rapv1 - 82.84
MARKET-1501 - (optional)
TSAI - (Any thing is okay as along as above is satisfied)
Works done :
####### Coding part starts ###########################
19-07-2020 - Attached the notebooks to run the file on pa100k and rapv2 dataset with baseline code
26-07-2020 - Added Vgg models code to run pa100k - (Vgg_valencebond_PA100k_notebook (1).ipynb)
30-07-2020 - Added Densenet models code to run pa100k - (Dense_net_PA100k_notebook.ipynb) - modified the input channel and batch size
30-07-2020 - Added Mobilenet models code to run pa100k - (Mobile_net_PA100k_notebook.ipynb) - modified input channel
30-07-2020 - Added MNAS net - (MNAS_net_PA100k_notebook.ipynb)
30-07-2020 - Added Squeeze net - (Squeze_net_PA100k_notebook.ipynb)
30-07-2020 - Added Alex net - (Alexnet_net_PA100k_notebook.ipynb)
30-07-2020 - Added Inception net - (Inception_net_PA100k_notebook.ipynb)
30-07-2020 - Added SE_resnet net - (se_resnet_net_PA100k_notebook.ipynb)
######### Coding part Completed ###########################
Authors of the base paper have not released original pkl file or dataset so we are going with existing methods
######### Hyper tunning and training part Start ###########################
Dataset link -
| Model name | Ckpt file | logs file | png file of final accuracy | epochs trained | number of trainval and test data |
|---|---|---|---|---|---|
| Resnet50 | drive | drive | ma: 0.7992, Acc: 0.7853, Prec: 0.8718, Rec: 0.8667 | 50 | trainval set: 90000, test set: 10000, attr_num : 26 |
| Densenet121 | drive | drive | ma: 0.7584, Acc: 0.7472, Prec: 0.8461, Rec: 0.8384 | 50 | trainval set: 90000, test set: 10000, attr_num : 26 |
| Alexnet | |||||
| mnasnet | |||||
| shufflenetv2 | |||||
| squeezenet | |||||
| vgg | |||||
| inception (optional if it works) | |||||
| seresnet |
Dataset link -
| Model name | Ckpt file | logs file | snap -png file of final accuracy | epochs trained | number of trainval and test data |
|---|---|---|---|---|---|
| Resnet50 | drive | drive | ma: 0.7862, Acc: 0.6640, Prec: 0.7794, Rec: 0.7974 | 50 | trainval set: 67943, test set: 16985, attr_num : 54 |
| Densenet121 | drive | drive | ma: 0.7666, Acc: 0.6488, Prec: 0.7729, Rec: 0.7801 | 50 | |
| trainval set: 67943, test set: 16985, attr_num : 54 | |||||
| Alexnet | drive | drive | ma: 0.7252, Acc: 0.6150, Prec: 0.7618, Rec: 0.7431, F1: 0.7468 | 50 | trainval set: 67943, test set: 16985, attr_num : 54 |
| mnasnet | |||||
| shufflenetv2 | |||||
| squeezenet | |||||
| vgg | drive | drive | ma: 0.7407, Acc: 0.6270, Prec: 0.7569, Rec: 0.7674, | 50 | trainval set: 67943, test set: 16985, attr_num : 54 |
| inception (optional if it works) | |||||
| seresnet |
Dataset link -
| Model name | Ckpt file | logs file | png file of final accuracy | epochs trained | number of trainval and test data |
|---|---|---|---|---|---|
| Resnet50 | drive | drive | ma: 0.8522, Acc: 0.7924 | 50 | trainset -11400, test set: 7600 |
| Densenet121 | drive | drive | - | 50 | trainset -11400, test set: 7600 |
| Alexnet | drive | drive | - | 50 | trainset -11400, test set: 7600 |
| mnasnet | drive | drive | - | 50 | trainset -11400, test set: 7600 |
| shufflenetv2 | drive | drive | - | 50 | trainset -11400, test set: 7600 |
| squeezenet | drive | drive | - | 50 | trainset -11400, test set: 7600 |
| vgg | drive | drive | - | 50 | trainset -11400, test set: 7600 |
| inception (optional if it works) | |||||
| se_resnet | drive | drive | - | 50 | trainset -11400, test set: 7600 |
Will update the accuracys soon
Resnet results - https://drive.google.com/drive/u/1/folders/1-FlKmYoj7wE0TsAgEjjFUqzAl_6-u6Eu
Dataset link -
| Model name | Ckpt file | logs file | snap -png file of final accuracy | epochs trained | number of trainval and test data |
|---|---|---|---|---|---|
| Resnet50 | |||||
| Densenet121 | |||||
| Alexnet | |||||
| mnasnet | |||||
| shufflenetv2 | |||||
| squeezenet | |||||
| vgg | |||||
| inception (optional if it works) | |||||
| seresnet |