IndexError while evaluate on COCO

When I run ssd300_evaluation_COCO.ipynb, an IndexError came out. I just modified the path and nothing more.

IndexError                                Traceback (most recent call last)
<ipython-input-11-2145451281d3> in <module>()
     11                     iou_threshold=0.45,
     12                     top_k=200,
---> 13                     normalize_coords=True)

~/ssd_keras/eval_utils/coco_utils.py in predict_all_to_json(out_file, model, img_height, img_width, classes_to_cats, data_generator, batch_size, data_generator_mode, model_mode, confidence_thresh, iou_threshold, top_k, pred_coords, normalize_coords)
    151     for i in tr:
    152         # Generate batch.
--> 153         batch_X, batch_image_ids, batch_inverse_transforms = next(generator)
    154         # Predict.
    155         y_pred = model.predict(batch_X)

~/ssd_keras/data_generator/object_detection_2d_data_generator.py in generate(self, batch_size, shuffle, transformations, label_encoder, returns, keep_images_without_gt, degenerate_box_handling)
    743 
    744                 # Check for degenerate ground truth bounding boxes before attempting any computations.
--> 745                 if np.any(batch_y[i][:,xmax] - batch_y[i][:,xmin] <= 0) or np.any(batch_y[i][:,ymax] - batch_y[i][:,ymin] <= 0):
    746                     if degenerate_box_handling == 'warn':
    747                         warnings.warn("Detected degenerate ground truth bounding boxes for batch item {} with bounding boxes {}, ".format(i, batch_y[i]) +

IndexError: list index out of range

[pierluigiferrari]

For inference we don’t want the uncertain predictions, but for the evaluation we do.

In order to reach high recall values in the computation of the precision-recall-curve you need predictions at various confidence levels well below 0.5.

[pierluigiferrari]

Thank you for reporting this, that’s indeed a bug. The generator tries to check for degenerate ground truth boxes even though there are no ground truth boxes.

I will push a fix later today. In the meantime you could fix this locally by adding the line if not (self.labels is None): just before the code block that checks for degenerate boxes.