search

liboyin / img-classify

Household Object Recognition Challenge
GNU General Public License v2.0
0 stars 0 forks source link

Libo's project timeline #7

Open liboyin opened 9 years ago

liboyin commented 9 years ago

Mar 27: Project proposal submitted. Apr 11: Pseudocode with SIFT and kNN. Apr 13: Prototype with SIFT and kNN on cosine distance. 80% accuracy. Apr 14: GitHub repo created. Apr 14: Tried custom kNN metric function: dist(a,b)=cos_dist(a,b) * abs(a.sift_scale - b.sift_scale). 50% accuracy. Apr 15: Migrated to scikit-learn framework. Experiment repeated 10 times on randomly train/test splits for stable result. Apr 20: Tried SVM with different kernels on default options. linear kernel gives 65% accuracy, polynomial kernel gives 83% accuracy, others lower than 5%. Apr 20: Tried random forest with both criterias on default options. Gini gives 58% accuracy, entropy gives 55% accuracy. Apr 20: Tried decision tree with both criterias on default options. Gini gives 52% accuracy, entropy gives 50% accuracy.

liboyin commented 9 years ago

Apr 24: Implemented new model as stated in https://github.com/liboyin/horc/issues/2. 91% accuracy.

liboyin commented 9 years ago

May 1: Refactored code for cleaner structure and easier debugging. May 1: Added error analysis of SNN model, as mentioned in https://github.com/liboyin/horc/issues/2. May 1: Confirmed the separability of classes in SNN model, as mentioned in https://github.com/liboyin/horc/issues/1.

liboyin commented 9 years ago

May 2: Migrated SNN code to probabilistic feamework. May 2: Implemented hue histogram intersection as global color classifier. Combined accuracy 95%.

liboyin commented 9 years ago

May 25: Moved code out of ipynb for better parallel support May 25: Re-wrote classifier for better parallel support May 25: Tried obtaining SIFT information from the hue channel. Not useful since some images only produce 3 sift key points.