CIMA experiments (#45)

* add CIMA pairing
* prepare CIMA experiments
* add notebook CIMA scope
* add notebook w. scope compare
* fix for too large images
* fix evaluation
* fix fig export
* add STD measures
* add JSON results
* update readm & add ref.
* update shell experiments
* drop enlighten
* minor rename

Author: Borda

.travis.yml

@@ -76,6 +76,7 @@ before_install:
     fi
 
 install:
+  - pip install "setuptools<46" -U  # v46 crashes openslide-python install
   - pip install -r requirements.txt
   - pip install -r ./tests/requirements.txt
   - pip --version ; pip list

README.md

@@ -320,6 +320,7 @@ The project is using the standard [BSD license](http://opensource.org/licenses/B
 
 For complete references see [bibtex](docs/references.bib).
 1. Borovec, J., Munoz-Barrutia, A., & Kybic, J. (2018). **[Benchmarking of image registration methods for differently stained histological slides](https://www.researchgate.net/publication/325019076_Benchmarking_of_image_registration_methods_for_differently_stained_histological_slides)**. In IEEE International Conference on Image Processing (ICIP) (pp. 3368–3372), Athens. [DOI: 10.1109/ICIP.2018.8451040](https://doi.org/10.1109/ICIP.2018.8451040)
+2. Borovec, J. (2019). BIRL: **Benchmark on Image Registration methods with Landmark validation**. arXiv preprint [arXiv:1912.13452.](https://arxiv.org/abs/1912.13452)
 
 ## Appendix - Useful information
 

appveyor.yml

@@ -62,6 +62,7 @@ install:
   # the parent CMD process).
   - "SET PATH=%PYTHON%;%PYTHON%\\Scripts;%PATH%"
   - python -m pip install --upgrade pip
+  - pip install "setuptools<46" -U  # v46 crashes openslide-python install
   - pip install -r requirements.txt
   - pip install -r ./tests/requirements.txt
   - pip install tox

birl/benchmark.py

@@ -455,6 +455,9 @@ def _perform_registration(self, df_row):
         row = self.__images_preprocessing(row)
         row[self.COL_TIME_PREPROC] = (time.time() - time_start) / 60.
         row = self._prepare_img_registration(row)
+        # if the pre-processing failed, return back None
+        if not row:
+            return None
 
         # measure execution time
         time_start = time.time()
@@ -468,6 +471,9 @@ def _perform_registration(self, df_row):
         row = self.__remove_pproc_images(row)
 
         row = self._parse_regist_results(row)
+        # if the post-processing failed, return back None
+        if not row:
+            return None
         row = self._clear_after_registration(row)
 
         if self.params.get('visual', False):

birl/utilities/drawing.py

@@ -324,11 +324,13 @@ def __init__(self, df, steps=5, fig=None, rect=None, fill_alpha=0.05, colors='ni
         for i, (idx, row) in enumerate(self.data.iterrows()):
             self.__draw_curve(idx, row, fill_alpha, color=colors[i], *args, **kwargs)
 
+        self._labels = []
         for ax in self.axes:
             for theta, label in zip(ax.get_xticks(), ax.get_xticklabels()):
                 self.__realign_polar_xtick(ax, theta, label)
+                self._labels.append(label)
 
-        self.ax.legend(loc='center left', bbox_to_anchor=(1.2, 0.7))
+        self._legend = self.ax.legend(loc='center left', bbox_to_anchor=(1.2, 0.7))
 
     @classmethod
     def __ax_set_invisible(self, ax):
@@ -490,7 +492,7 @@ def draw_matrix_user_ranking(df_stat, higher_better=False, fig=None, cmap='tab20
     ranking = compute_matrix_user_ranking(df_stat, higher_better)
 
     if fig is None:
-        fig, _ = plt.subplots(figsize=np.array(df_stat.as_matrix().shape[::-1]) * 0.35)
+        fig, _ = plt.subplots(figsize=np.array(df_stat.values.shape[::-1]) * 0.35)
     ax = fig.gca()
     arange = np.linspace(-0.5, len(df_stat) - 0.5, len(df_stat) + 1)
     norm = plt_colors.BoundaryNorm(arange, len(df_stat))
@@ -513,7 +515,7 @@ def draw_scatter_double_scale(df, colors='nipy_spectral',
                               figsize=None,
                               legend_style=None,
                               plot_style=None,
-                              x_spread=(0.3, 5)):
+                              x_spread=(0.4, 5)):
     """Draw a scatter with double scales on left and right
 
     :param DF df: dataframe
@@ -531,7 +533,7 @@ def draw_scatter_double_scale(df, colors='nipy_spectral',
     >>> df = pd.DataFrame(np.random.random((10, 3)), columns=['col1', 'col2', 'col3'])
     >>> fig, axs = draw_scatter_double_scale(df, ax_decs={'name': None}, xlabel='X')
     >>> axs  # doctest: +ELLIPSIS
-    (<...>, None)
+    {...}
     >>> # just the selected columns
     >>> fig, axs = draw_scatter_double_scale(df, ax_decs={'name1': ['col1', 'col2'],
     ...                                                   'name2': ['col3']})
@@ -602,5 +604,7 @@ def draw_scatter_double_scale(df, colors='nipy_spectral',
     # legend - https://matplotlib.org/3.1.1/gallery/text_labels_and_annotations/custom_legends.html
     if legend_style is None:
         legend_style = dict(loc='upper center', bbox_to_anchor=(1.25, 1.0), ncol=1)
-    ax1.legend(idx_names, **legend_style)
-    return fig, (ax1, ax2)
+    lgd = ax1.legend(idx_names, **legend_style)
+
+    extras = {'ax1': ax1, 'ax2': ax2, 'legend': lgd}
+    return fig, extras

birl/utilities/evaluate.py

@@ -219,7 +219,7 @@ def compute_matrix_user_ranking(df_stat, higher_better=False):
            [ 0.,  2.,  1.],
            [ 4.,  4.,  2.]])
     """
-    ranking = np.zeros(df_stat.as_matrix().shape)
+    ranking = np.zeros(df_stat.values.shape)
     nan = -np.inf if higher_better else np.inf
     for i, col in enumerate(df_stat.columns):
         vals = [v if not np.isnan(v) else nan for v in df_stat[col]]

bm_ANHIR/Dockerfile

@@ -47,7 +47,7 @@ ENV PATH="/home/evaluator/.local/bin:${PATH}"
 COPY --chown=evaluator:evaluator ./evaluate_submission.py /opt/evaluation/
 COPY --chown=evaluator:evaluator ./dataset_ANHIR/dataset_medium.csv /opt/evaluation/dataset.csv
 COPY --chown=evaluator:evaluator ./dataset_ANHIR/computer-performances_cmpgrid-71.json /opt/evaluation/computer-performances.json
-COPY --chown=evaluator:evaluator ./dataset_ANHIR/landmarks_user /opt/evaluation/lnds_provided
+COPY --chown=evaluator:evaluator dataset_ANHIR/landmarks_user_phase1 /opt/evaluation/lnds_provided
 COPY --chown=evaluator:evaluator ./dataset_ANHIR/landmarks_all /opt/evaluation/lnds_reference
 
 # Define execution

bm_ANHIR/evaluate_submission.py

@@ -209,18 +209,25 @@ def parse_landmarks(idx_row):
         # 'reference landmarks': np.round(lnds_ref, 1).tolist(),
         # 'warped landmarks': np.round(lnds_warp, 1).tolist(),
         'matched-landmarks': match_lnds,
-        'Robustness': row.get(ImRegBenchmark.COL_ROBUSTNESS, 0),
-        'Norm-Time_minutes': row.get(COL_NORM_TIME, None),
+        'Robustness': np.round(row.get(ImRegBenchmark.COL_ROBUSTNESS, 0), 3),
+        'Norm-Time_minutes': np.round(row.get(COL_NORM_TIME, None), 5),
         'Status': row.get(ImRegBenchmark.COL_STATUS, None),
     }
+
+    def _round_val(row, col):
+        dec = 5 if col.startswith('rTRE') else 2
+        return np.round(row[col], dec)
+
     # copy all columns with Affine statistic
-    item.update({col.replace(' ', '-'): row[col] for col in row if 'affine' in col.lower()})
+    item.update({col.replace(' ', '-'): _round_val(row, col)
+                 for col in row if 'affine' in col.lower()})
     # copy all columns with rTRE, TRE and Overlap
     # item.update({col.replace(' (final)', '').replace(' ', '-'): row[col]
     #              for col in row if '(final)' in col})
-    item.update({col.replace(' (elastic)', '_elastic').replace(' ', '-'): row[col]
+    item.update({col.replace(' (elastic)', '_elastic').replace(' ', '-'): _round_val(row, col)
                  for col in row if 'TRE' in col})
-    return idx, item
+    # later in JSON keys ahs to be str only
+    return str(idx), item
 
 
 def compute_scores(df_experiments, min_landmarks=1.):
@@ -271,6 +278,7 @@ def _compute_scores_general(df_experiments, df_expt_robust):
     # parse specific metrics
     scores = {
         'Average-Robustness': np.mean(df_experiments[ImRegBenchmark.COL_ROBUSTNESS]),
+        'STD-Robustness': np.std(df_experiments[ImRegBenchmark.COL_ROBUSTNESS]),
         'Median-Robustness': np.median(df_experiments[ImRegBenchmark.COL_ROBUSTNESS]),
         'Average-Rank-Median-rTRE': np.nan,
         'Average-Rank-Max-rTRE': np.nan,
@@ -280,15 +288,18 @@ def _compute_scores_general(df_experiments, df_expt_robust):
                       ('Max-rTRE', 'rTRE Max'),
                       ('Average-rTRE', 'rTRE Mean'),
                       ('Norm-Time', COL_NORM_TIME)]:
-        scores['Average-' + name] = np.nanmean(df_experiments[col])
-        scores['Average-' + name + '-Robust'] = np.nanmean(df_expt_robust[col])
-        scores['Median-' + name] = np.median(df_experiments[col])
-        scores['Median-' + name + '-Robust'] = np.median(df_expt_robust[col])
+        for df, sufix in [(df_experiments, ''), (df_expt_robust, '-Robust')]:
+            scores['Average-' + name + sufix] = np.nanmean(df[col])
+            scores['STD-' + name + sufix] = np.nanstd(df[col])
+            scores['Median-' + name + sufix] = np.median(df[col])
     return scores
 
 
 def _compute_scores_state_tissue(df_experiments):
     scores = {}
+    if ImRegBenchmark.COL_STATUS not in df_experiments.columns:
+        logging.warning('experiments (table) is missing "%s" column', ImRegBenchmark.COL_STATUS)
+        df_experiments[ImRegBenchmark.COL_STATUS] = 'any'
     # filter all statuses in the experiments
     statuses = df_experiments[ImRegBenchmark.COL_STATUS].unique()
     # parse metrics according to TEST and TRAIN case

bm_CIMA/README.md

@@ -0,0 +1,32 @@
+# Experimentation CIMA dataset
+
+This section is strictly limited to image registration experiment on [CIMA dataset](http://cmp.felk.cvut.cz/~borovji3/?page=dataset).
+
+## Structure
+
+- **Datasets**: the particular dataset setting described in the image/landmarks pairing - csv tables called `dataset_CIMA_<scope>.csv`
+- **Script**: the execution script is `run-SOTA-experiments.sh` and it perform all experiments
+- **Results**: the experimental results are exported and zipped per particular dataset scope, the archives are `results_size-<scope>.zip`
+
+
+## Usage
+
+**Reproduce statistic**
+
+You need to unzip the particular result for each dataset scale in to a separate folder (e.g with the same name).
+Then  you need to run the [scope notebook](../notebooks/CIMA_SOTA-results_scope.ipynb) for showing results on a particular dataset scope or [comparing notebook](../notebooks/CIMA_SOTA-results_comparing.ipynb) to compare some statistics over two scopes.
+Note that with using attached JSON results you do not need to run cells related parsing results from raw benchmarks results.
+
+**Add own method to statistic**
+
+You need to run your benchmark on the particular dataset scope, the image oaring are:
+- [10k scope](dataset_CIMA_10k.csv)
+- [full scope](dataset_CIMA_full.csv)
+
+Then you can parse just the new results with [evaluation script](../bm_ANHIR/evaluate_submission.py) or execute the parsing cells at the beginning of [scope notebook](../notebooks/CIMA_SOTA-results_scope.ipynb).
+
+
+## References
+
+For complete references see [bibtex](../docs/references.bib).
+1. Borovec, J. (2019). **BIRL: Benchmark on Image Registration methods with Landmark validation**. arXiv preprint [arXiv:1912.13452.](https://arxiv.org/abs/1912.13452)