# Internship Report - Name : Harsh Kumar - Duration : 6th June - 31st July - username : ufh2kor In these two months at BGSW, I learned and worked under the guidance of Mr. Vinoj. My work consisted of majorly two projects 1. Python developement on Internal LabelConvertor 2. Research on the general problem of Image Similarity Search ## Internal Label Convertor Contributing to the [Internal_LabelConvertor](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/browse) repository, my work was focused on writing label convertors and unit tests for them. Major PRs : [PR#15](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/pull-requests/15/overview), [PR#12](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/pull-requests/12/overview), [PR#13](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/pull-requests/13/overview) ### Worked on coco labels - Developed `coco.py` with read and write methods to for the [CoCo](https://cocodataset.org/) format - Wrote convertors to and from `Laas` in `coco_to_laas.py` and `laas_to_coco.py` - Wrote unit tests for these convertors - Major commits : [PR#12](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/pull-requests/12/overview), [#f6988c2de12](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/pull-requests/12/commits/f6988c2de12cd6a5786c319f432c65be38fa485e) ### Initiated a unitest developement enviournment in the repository - Studied the codebase and understood previous work to write better test cases - test config files for convertors in `test_config_files` - Developed `test_bas_to_pixel_label.py` - Later followed suit to develope test cases for all the convertors that I wrote - Major commits : [#0c74d3937af](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/pull-requests/12/commits/0c74d3937af7decf2e75725c116fb3169f689b56), [#b3a788d62be](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/commits/b3a788d62bee246aaa6b9cf48bb9b61f72ed4767) [#ca61d978b2f](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/pull-requests/12/commits/ca61d978b2f4d749f719727066ecbf1245e23c8e) ### Worked on Laas to PixelLabel - Convertor to and fro from these labels - Wrote tests cases for both of them - Major commits : [#55bdef0eba3](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/commits/55bdef0eba34bb647e087e85610b9b4f904006af), [#b3a788d62be](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/commits/b3a788d62bee246aaa6b9cf48bb9b61f72ed4767) ### PixelLabel to Semseg - Convertor to and fro from these labels - Ongoing : [PR#15](https://sourcecode.socialcoding.bosch.com/projects/ERD/repos/internal_labelconvertor/pull-requests/15/overview) ---- ## Image Similarity Search Problem Reserached for ideas on how to solve the Image Similarity Search Problem for a ongoing project at BGSW. The niche we were interested in was to apply these techniques to find a given patch from a test image in a big _unlabelled_ dataset. I approached this problem by going into depth as well width on the availble litreture. After reading a number of papers and blogs on this topic, I presented my ideas in [this](https://docs.google.com/presentation/d/1wAB37q0hW6Y5Qfz5sY9cZwgJjKuUQsmOWMh0wMF9U88/edit?usp=sharing) presentation. I shall try to summarise my work in the following section. #### Broadally I came up with these ideas : - Contrastive Learning - Siamese Networks - Deep Hashing - AutoEncoders - GPU Optimisation Most promising idea among these was `Contrastive Learning` . Upon discussion we decided to go forward with the `MoCo` framework. I then set up the model and am currently in the process of using this model on a [toy dataset](https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/) to check model's feasability ----- Finally, I cordially thank my manager and my co-interns at BGSW for their help and constant guidance.