Skip to content

NOTES.md

Latest commit

 

History

History
89 lines (75 loc) · 2.96 KB

NOTES.md

File metadata and controls

89 lines (75 loc) · 2.96 KB

0.0.6.dev

Package

  • Drop cv2 dependence (just numpy, skimage, scipy) (still used somewhere, but not necessary anymore)

Training

  • DTree weak classifier
  • Fern classifier CANCELED
  • Agnostic to underlying feature type
  • Quantization
    • quantized thresholds
    • quantized response tables (hs)
  • Feature banks (fpga support)
  • fit_stage must return training history
  • Rejection scheduling in fit_stage
  • Degenerated solutions - training dstump on 0 samples
    • 1/ avoid if possible
    • 2/ solve when it happens
  • Check what happens with weights
  • Model saving and restoration (h5) Protobuf

Channels

  • Configurable channels
  • Integer channels
  • More than one channel in image/sample

Detector

  • Support for dtree and dstump

Sampling and ground truth

  • bb distance must take sizes and aspect ratios into account
  • Random image adjustments
  • rewrite sample SamplePool
  • bbGt reader parameters
  • Ignore bounding boxes
  • Stop updating pool at some point, e.g. P0 < 1e-7 or so
  • Sample pool without generator???
  • Sample pool should accept model in constructor (is mutable and can be used in update method without explicitly passing it)
  • Pool should not hold prior probs. should be in training module

0.1.0.dev

Package

  • Update Doc
  • Input parameter checking - ValueError, TypeError where appropriate
  • Example data and notebook
  • Logging to screen and to file
  • Applications
    • training - waldboost-train -i dataset --n0=1000 --alpha=0.2 -o test.pb
    • detection - waldboost-detect -m test.pb image.jpg image1.jpg ...
  • Handle corner cases - no boxes on image, image too small for channels etc.

Training

  • Better API for Learner - methods object_prob, bg_prob, loss
  • Resume learning
    • save/load learner to file
    • supply learer to training
  • Constant soft cascade pipeline

Model and channels

  • Channel parametrs in proto
  • Discard old messages from proto and simplify structure
  • Correct symbol initialization when loading
  • classifier as list of DTree, theta in its own array so it can be modified
  • Channel opts as object
  • Optimizations
    • Numba and CUDA for one selected variant
  • Stats
    • Keep worst case eval. cost (possibly rewrite stats gthering)

Data sampling

  • Change image set during training Fixed by "resume learning"
  • Either update bbx package to be more consistent or use bbox (or other) package
  • Better API for Pool
  • Fix dt_boxes is None and gt_boxes is None cases
  • Change data generator to yield dict with "image", "ground_truth_boxes" keys

Ground Truth and Bounding Box

  • Improve bbGt reading - return BoxList
  • Utility functions in bbox module - empty, empty_like, scale, set_aspect_ratio, ....

Verification

  • Consider rename module - postprocessing?
  • Verification - better log, validation set, show fp/fn rates
  • bbox regression?

FIXME

  • SamplePool is slow when len(self.samples) gets large