- Documentation
- Add links to papers.
- Add links to project files.
- Make new notebook showing differences in cropped images that result from jpeg compression.
- Organization and Refactoring
- Move all plotting functions to visualize.py
-
Move create_label_metrics into create_image_metrics?Keep separate. - Don't use multiple folders for jupyter notebooks. Just makes it hard to find.
-
Create a new SDCTile class that represents a single tile from SDCImage.Cancelled. Too slow.
- Performance
- Parallelize the creation of image_metrics.
- Parallelize the creation of overlap_scores.
- Write unit tests
- ...for most everything in utils.py
- Use design pattern: Each test should have 3 parts: Arrange, Act, Assert
- Verify rotate image {1, 2, 3} x 90 degrees in pytorch?
- make sure
len(md5hashes) == len(set(md5hashes)), otherwise we'll have to make our hashes longer.
- Net upgrades
- Try using BatchNorm, or LayerNorm
- Try an all convolutional net. Remove max pooling layers and add stride 2 (or 3?) to convolutions
- Try using inputs with shape (BATCH_SIZE, 2, 3, 256, 256) instead of (BATCH_SIZE, 6, 256, 256)?
- Importance sampling
- Implement UCB like function?
- Implement TF-IDF like function?
- Use bmh hamming distance with as_scores=True as initial guess to weights?
- Features
- Implement an efficient way to lower (or raise) the overlap matches threshold.
- Implement efficient way to remove truth from auto when they are added manually.
- Skip image pairs that have already been verified as duplicates.
- Find out which remaining tile pairs have masks but aren't in dup_truth.
- Provide examples of false positives and false negatives for each of the following:
- image hashes
- image entropy
- image histograms
- grey-level co-occurrence matrix
- ship overlap
- Use decision trees to show performance of DupNet over other metrics.
- Dup Tool
- Upgrade the dup_tool to use Bokeh or Plotly for interactions and rendering.
- Handle StopIteration exception on last record read.