[OD][bug] models' loss doesn't use is_crowd

[netanel-s]

System information

• What is the top-level directory of the model you are using: object_detection
• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): no
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): NA
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): NA
• Bazel version (if compiling from source): NA
• CUDA/cuDNN version: NA
• GPU model and memory: NA
• Exact command to reproduce: NA

Describe the problem

It seems that when training Faster R-CNN or SSD model, their loss doesn't take into account is_crowd annotations (see RPN and classifier in Faster R-CNN, SSD). This means two problems:

1. The single bounding box of the entire crowd is used as a positive example.
2. The crowded objects inside the annotation will be assigned as negative examples, when naively they should be set as ignored examples.

These two problems will cause all models to be suboptimal or even bad at detecting objects in crowds.

[ppwwyyxx]

My implementation of Faster/Mask R-CNN handles crowd annotations. It does not use them as positive examples. And it can label anchors close to crowd region as ignored