- Investigating the Ability of CNNs to Count Visual Concepts
- Fergus Steel
- 2542391s
- Dr. Paul Siebert
- This file contains the time log for your project. It will be submitted along with your final dissertation.
- YOU MUST KEEP THIS UP TO DATE AND UNDER VERSION CONTROL.
- This timelog should be filled out honestly, regularly (daily) and accurately. It is for your benefit.
- Follow the structure provided, grouping time by weeks. Quantise time to the half hour.
- 0.5 hours, made initial contact with supervisor.
- 1 hour Set up version control system, and customised template.
- 1 hour Created meeting minutes template, and planned for meeting01 with supervisor.
- 0.5 hour Filled out the "summer" document.
- 1 hour Finished "Counting and locating high-density objects using convolutional neural networks" and continued to develop ideas and plan for meeting.
- 0.5 hours Researched relevant architectures (R-CNN, Fast R-CNN, YOLO)
- 0.5 hours Read the RetinaNet paper and made notes.
- 1.5 hours Read the survey of crowd counting techniques paper.
- 1 hour Read the DPSB paper + notes
- 2 hour Initial Meeting with Paul, Introductions and Advice.
- 0.5 hours Read Feature Pyramid Network Paper to develop understanding of feature extractors.
- 2 hour Panoptic Segmentation Research, OneFormer, Mask2Former, K-Net and LPSNet investigations began reading papers.
- 0.5 hours Began considering focussing on Attention Mechanisms and the role they may be able to perform?
- 1 hour Attempting to install tf-models-official for the purpose of demonstration of models but encountered dependency clashes and had to reinstall several times, should really be using a virtual environment...
- 0.5 hours Initial plan for Semester 1
- 0.5 hours Read another crowd counting article - https://www.analyticsvidhya.com/blog/2021/06/crowd-counting-using-deep-learning/
- 1 hour Read Convolutional Neural Network Chapter in "Artificial Intelligence, a Modern Approach"
- 0.5 hours Read NARX paper (Lin et al. 1996)
- 0.5 hours Read CNN-NARX paper (Jalal et al. 2019)
- 1 hours Read MIST-RNN paper, did further NARX research.
- 0.5 hours further temporal feature research (ConvLSTM);
- 2 hours Tried to fix NVIDIA Graphics Drivers on laptop (failure)
- 1 hour drafted research direction ideas.
- 3 hours Reading existing Spatio-Temporal papers, understanding techniques used.
- 1.5 hours Further NARX reserach, read predicting Chaotic Time Series which goes into muhc more detail on the artchitectures of NARX networks
- 0.5 hours Meeting 2 Prep
- 2 hours Supervisor Meeting 2
- 1.16 hour Watched Geoffrey Hinton's "What is wrong with convoultional neural nets?"
- 4 hours Spent all day researching then implementing a demonstration before I bricked my laptop and lost all progress from this day
- 0.5 hours watched Capsule networks - A survey by Dr. Yogesh Rawat, UCF
- 3 hours Few-shot / Zero Shot Object counting research. - can we use capsule networks in exempler inspection?
- 1 hour further Zero Shot Research
- 2 hours read more current literature on FSC (general class agnostic counting research)
- 1 hour Planning for meeting 3
- 2 hours planning and research for meeting 3
- 2.5 hours Third Supervisor meeting
- 0.5 hours Wrote up meeting minutes for third supervisor meeting.
- 1.5 hours Read some more Class Agnostic Counting papers thanks to connected papers graph
- 1.5 hours Read some more Capsule Network papers and discovered crowding problem.
- 1 hour Started proposal
- 1 hour Meeting 4 Preperation
- 2 hours Meeting 4
- 1 hour further proposal work
- 0.5 hours writing up meeting 4 notes
- 1 hour finishing proposal.md document
- 0.5 hour Read Joseph William's writing style essay before dissertation writing.
- 1 hour draft reserach questions and collect opinions
- 0.5 hours start to draft experimental plan (Hypothesis and Method)
- 0.5 hours Revisited FSC Objecting Counting Research (You et al. (2022))
- 1 hour Dissertation outline (Headings and further planning)
- 1 hour Research Design Writing
- 1 hour fixing bibtex, begun writing dissertation
- 1 hour dataset decisions + alt. plan
- 1 hours adapting proposal into dissertation (unfinsihed)
- 1 hour Meeting 5 Plan
- 1 hour Network Planning (More CapsNet Research)
- 0.5 hours Studied Keras implementation of CapsNet https://github.com/XifengGuo/CapsNet-Keras/blob/master
- 0.75 hours slides for meeting 5
- 2.5 hours Meeting 5
- 2 hours Dataset Planning + Demo Implementation Start
- 3 hours Demo Dataset Generator Almost Completed
- 2 hours Finished Dataset Generator - could still add rotation and some usability features but needs to be adapted to python files anyway
- 0.5 hours read Matrix Capsuyle Em routing appaer
- 1 hour "Counting" reserach, reading about counting in organisms.
- 1 hour Meeting 6 Prep, planning pilot study.
- 0.5 hours Meeting 6 Prep
- 2 hours Conversion to .npy files (this is a failure, because I used wrong data type and need to cast it to a smaller datatype and ensure that it stays as fine grained as I need it to)
- 1 hour began coding pilot study, using CNNs in order to counter and contrast.
- 2.5 hours Meeting 7
- 2 hours pilot study coding, data preprocessing and initial model design
- 4 hours Remade pilot study, using UNET architecutre
- 3 hours Dissertation writing, Capsule Network background
- 3 hours Pilot training, tweaking, why wont this work!
- 2 hours Pilot training debugged, its working suuiiiiiii! (now just need to tweak it)
- 3 hours Remaking data set to include segmentation masks rather than gaussians
- 3 hours Retraining, tweaking, trying to get new mask network working
- 2 hours started on cleaning repository and making code more readable
- 3 hours Visual system research, mainly dissertation readings
- NO WORK DONE, Exam break
- 5 hours Adapting loss function to incorporate Multiscale Structural Simularity of each density map.
- 1.5 hours Wrote halfway report, planned out semester 2 work
- 2 hours Masking research, how to investigate capsule outputs?
- 2 hours back to school, getting back into gear and figuring out where I am in the project
- 3 hours tweaking code to optimise the best-case performance.
- 1 hour Panicking, trying to fix the "best-case" performance which is really not good enough
- 3 Hours Making Network Larger and testing if this helps capture more complexity in outputs
- 1 hour Started working onevalluation metrics, even if its not working I still need somethign to graph
- 4 hours Added disjoint loss function to stop cautious guess, needs tweaking
- 1 hour Testing the effect of increasing model complexity
- 0.5 hours Bug fixing, kernel size 1 bug
- 1 hour quick meeting with Paul to discuss dissertation tips
- 3 horus Rough draft of introduction
- 0.5 hours collating background info into background
- ALL DAY Reading through previous dissertation attempts of this subject
- 2 hours Exploring using simplified version of model for better results
- 1 Hour Finished off first draft of the skeleton of the dissertation now I have some direction
- 1 hour Finally started looking into visualizing capsules (made a python file, barely got started)
- 1 hour Writing some background on CNN reserach
- 3 hours Exploring previous version of model, fixing bugs, testing
- 0.5 Hours Backdating timelog with all the panic that has occured over last 3 weeks