y2020/vision/camera_reader.cc

#include <opencv2/calib3d.hpp>
#include <opencv2/features2d.hpp>
#include <opencv2/imgproc.hpp>

#include "aos/events/shm_event_loop.h"
#include "aos/flatbuffer_merge.h"
#include "aos/init.h"
#include "aos/network/team_number.h"
#include "y2020/vision/sift/sift971.h"
#include "y2020/vision/sift/sift_generated.h"
#include "y2020/vision/sift/sift_training_generated.h"
#include "y2020/vision/tools/python_code/sift_training_data.h"
#include "y2020/vision/v4l2_reader.h"
#include "y2020/vision/vision_generated.h"

// config used to allow running camera_reader independently.  E.g.,
// bazel run //y2020/vision:camera_reader -- --config y2020/config.json
//   --override_hostname pi-7971-1  --ignore_timestamps true
DEFINE_string(config, "config.json", "Path to the config file to use.");
DEFINE_bool(skip_sift, false,
            "If true don't run any feature extraction.  Just forward images.");
DEFINE_bool(ransac_pose, false,
            "If true, do pose estimate with RANSAC; else, use ITERATIVE mode.");
DEFINE_bool(use_prev_pose, true,
            "If true, use previous pose estimate as seed for next estimate.");

namespace frc971 {
namespace vision {
namespace {

class CameraReader {
 public:
  CameraReader(aos::EventLoop *event_loop,
               const sift::TrainingData *training_data, V4L2Reader *reader,
               const cv::Ptr<cv::flann::IndexParams> &index_params,
               const cv::Ptr<cv::flann::SearchParams> &search_params)
      : event_loop_(event_loop),
        training_data_(training_data),
        camera_calibration_(FindCameraCalibration()),
        reader_(reader),
        image_sender_(event_loop->MakeSender<CameraImage>("/camera")),
        result_sender_(
            event_loop->MakeSender<sift::ImageMatchResult>("/camera")),
        detailed_result_sender_(
            event_loop->MakeSender<sift::ImageMatchResult>("/camera/detailed")),
        read_image_timer_(event_loop->AddTimer([this]() { ReadImage(); })),
        prev_R_camera_field_vec_(cv::Mat::zeros(3, 1, CV_32F)),
        prev_T_camera_field_(cv::Mat::zeros(3, 1, CV_32F)) {

    for (int ii = 0; ii < number_training_images(); ++ii) {
      matchers_.push_back(cv::FlannBasedMatcher(index_params, search_params));
    }

    CopyTrainingFeatures();

    for (auto &matcher : matchers_) {
      matcher.train();
    }

    event_loop->OnRun(
        [this]() { read_image_timer_->Setup(event_loop_->monotonic_now()); });
  }

 private:
  const sift::CameraCalibration *FindCameraCalibration() const;

  // Copies the information from training_data_ into matcher_.
  void CopyTrainingFeatures();
  // Processes an image (including sending the results).
  void ProcessImage(const CameraImage &image);
  // Reads an image, and then performs all of our processing on it.
  void ReadImage();

  flatbuffers::Offset<
      flatbuffers::Vector<flatbuffers::Offset<sift::ImageMatch>>>
  PackImageMatches(flatbuffers::FlatBufferBuilder *fbb,
                   const std::vector<std::vector<cv::DMatch>> &matches);
  flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<sift::Feature>>>
  PackFeatures(flatbuffers::FlatBufferBuilder *fbb,
               const std::vector<cv::KeyPoint> &keypoints,
               const cv::Mat &descriptors);

  void SendImageMatchResult(const CameraImage &image,
                            const std::vector<cv::KeyPoint> &keypoints,
                            const cv::Mat &descriptors,
                            const std::vector<std::vector<cv::DMatch>> &matches,
                            const std::vector<cv::Mat> &camera_target_list,
                            const std::vector<cv::Mat> &field_camera_list,
                            const std::vector<cv::Point2f> &target_point_vector,
                            const std::vector<float> &target_radius_vector,
                            const std::vector<int> &training_image_indices,
                            aos::Sender<sift::ImageMatchResult> *result_sender,
                            bool send_details);

  // Returns the 2D (image) location for the specified training feature.
  cv::Point2f Training2dPoint(int training_image_index,
                              int feature_index) const {
    const float x = training_data_->images()
                        ->Get(training_image_index)
                        ->features()
                        ->Get(feature_index)
                        ->x();
    const float y = training_data_->images()
                        ->Get(training_image_index)
                        ->features()
                        ->Get(feature_index)
                        ->y();
    return cv::Point2f(x, y);
  }

  // Returns the 3D location for the specified training feature.
  cv::Point3f Training3dPoint(int training_image_index,
                              int feature_index) const {
    const sift::KeypointFieldLocation *const location =
        training_data_->images()
            ->Get(training_image_index)
            ->features()
            ->Get(feature_index)
            ->field_location();
    return cv::Point3f(location->x(), location->y(), location->z());
  }

  const sift::TransformationMatrix *FieldToTarget(int training_image_index) {
    return training_data_->images()
        ->Get(training_image_index)
        ->field_to_target();
  }

  void TargetLocation(int training_image_index, cv::Point2f &target_location,
                      float &target_radius) {
    target_location.x =
        training_data_->images()->Get(training_image_index)->target_point_x();
    target_location.y =
        training_data_->images()->Get(training_image_index)->target_point_y();
    target_radius = training_data_->images()
                        ->Get(training_image_index)
                        ->target_point_radius();
  }

  int number_training_images() const {
    return training_data_->images()->size();
  }

  cv::Mat CameraIntrinsics() const {
    const cv::Mat result(3, 3, CV_32F,
                         const_cast<void *>(static_cast<const void *>(
                             camera_calibration_->intrinsics()->data())));
    CHECK_EQ(result.total(), camera_calibration_->intrinsics()->size());
    return result;
  }

  cv::Mat CameraDistCoeffs() const {
    const cv::Mat result(5, 1, CV_32F,
                         const_cast<void *>(static_cast<const void *>(
                             camera_calibration_->dist_coeffs()->data())));
    CHECK_EQ(result.total(), camera_calibration_->dist_coeffs()->size());
    return result;
  }

  aos::EventLoop *const event_loop_;
  const sift::TrainingData *const training_data_;
  const sift::CameraCalibration *const camera_calibration_;
  V4L2Reader *const reader_;
  std::vector<cv::FlannBasedMatcher> matchers_;
  aos::Sender<CameraImage> image_sender_;
  aos::Sender<sift::ImageMatchResult> result_sender_;
  aos::Sender<sift::ImageMatchResult> detailed_result_sender_;
  // We schedule this immediately to read an image. Having it on a timer means
  // other things can run on the event loop in between.
  aos::TimerHandler *const read_image_timer_;

  // Storage for when we want to use the previous estimates of pose
  cv::Mat prev_R_camera_field_vec_;
  cv::Mat prev_T_camera_field_;

  const std::unique_ptr<frc971::vision::SIFT971_Impl> sift_{
      new frc971::vision::SIFT971_Impl()};
};

const sift::CameraCalibration *CameraReader::FindCameraCalibration() const {
  const std::string_view node_name = event_loop_->node()->name()->string_view();
  const int team_number = aos::network::GetTeamNumber();
  for (const sift::CameraCalibration *candidate :
       *training_data_->camera_calibrations()) {
    if (candidate->node_name()->string_view() != node_name) {
      continue;
    }
    if (candidate->team_number() != team_number) {
      continue;
    }
    return candidate;
  }
  LOG(FATAL) << ": Failed to find camera calibration for " << node_name
             << " on " << team_number;
}

void CameraReader::CopyTrainingFeatures() {
  int training_image_index = 0;
  for (const sift::TrainingImage *training_image : *training_data_->images()) {
    cv::Mat features(training_image->features()->size(), 128, CV_32F);
    for (size_t i = 0; i < training_image->features()->size(); ++i) {
      const sift::Feature *feature_table = training_image->features()->Get(i);

      // We don't need this information right now, but make sure it's here to
      // avoid crashes that only occur when specific features are matched.
      CHECK(feature_table->has_field_location());

      const flatbuffers::Vector<uint8_t> *const descriptor =
          feature_table->descriptor();
      CHECK_EQ(descriptor->size(), 128u) << ": Unsupported feature size";
      const auto in_mat = cv::Mat(
          1, descriptor->size(), CV_8U,
          const_cast<void *>(static_cast<const void *>(descriptor->data())));
      const auto out_mat = features(cv::Range(i, i + 1), cv::Range(0, 128));
      in_mat.convertTo(out_mat, CV_32F);
    }
    matchers_[training_image_index].add(features);
    ++training_image_index;
  }
}

void CameraReader::SendImageMatchResult(
    const CameraImage &image, const std::vector<cv::KeyPoint> &keypoints,
    const cv::Mat &descriptors,
    const std::vector<std::vector<cv::DMatch>> &matches,
    const std::vector<cv::Mat> &camera_target_list,
    const std::vector<cv::Mat> &field_camera_list,
    const std::vector<cv::Point2f> &target_point_vector,
    const std::vector<float> &target_radius_vector,
    const std::vector<int> &training_image_indices,
    aos::Sender<sift::ImageMatchResult> *result_sender, bool send_details) {
  auto builder = result_sender->MakeBuilder();
  const auto camera_calibration_offset =
      aos::RecursiveCopyFlatBuffer(camera_calibration_, builder.fbb());

  flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<sift::Feature>>>
      features_offset;
  flatbuffers::Offset<
      flatbuffers::Vector<flatbuffers::Offset<sift::ImageMatch>>>
      image_matches_offset;
  if (send_details) {
    features_offset = PackFeatures(builder.fbb(), keypoints, descriptors);
    image_matches_offset = PackImageMatches(builder.fbb(), matches);
  }

  std::vector<flatbuffers::Offset<sift::CameraPose>> camera_poses;

  CHECK_EQ(camera_target_list.size(), field_camera_list.size());
  for (size_t i = 0; i < camera_target_list.size(); ++i) {
    cv::Mat camera_target = camera_target_list[i];
    CHECK(camera_target.isContinuous());
    const auto data_offset = builder.fbb()->CreateVector<float>(
        reinterpret_cast<float *>(camera_target.data), camera_target.total());
    const flatbuffers::Offset<sift::TransformationMatrix> transform_offset =
        sift::CreateTransformationMatrix(*builder.fbb(), data_offset);

    cv::Mat field_camera = field_camera_list[i];
    CHECK(field_camera.isContinuous());
    const auto fc_data_offset = builder.fbb()->CreateVector<float>(
        reinterpret_cast<float *>(field_camera.data), field_camera.total());
    const flatbuffers::Offset<sift::TransformationMatrix> fc_transform_offset =
        sift::CreateTransformationMatrix(*builder.fbb(), fc_data_offset);

    const flatbuffers::Offset<sift::TransformationMatrix>
        field_to_target_offset = aos::RecursiveCopyFlatBuffer(
            FieldToTarget(training_image_indices[i]), builder.fbb());

    sift::CameraPose::Builder pose_builder(*builder.fbb());
    pose_builder.add_camera_to_target(transform_offset);
    pose_builder.add_field_to_camera(fc_transform_offset);
    pose_builder.add_field_to_target(field_to_target_offset);
    pose_builder.add_query_target_point_x(target_point_vector[i].x);
    pose_builder.add_query_target_point_y(target_point_vector[i].y);
    pose_builder.add_query_target_point_radius(target_radius_vector[i]);
    camera_poses.emplace_back(pose_builder.Finish());
  }
  const auto camera_poses_offset = builder.fbb()->CreateVector(camera_poses);

  sift::ImageMatchResult::Builder result_builder(*builder.fbb());
  result_builder.add_camera_poses(camera_poses_offset);
  if (send_details) {
    result_builder.add_image_matches(image_matches_offset);
    result_builder.add_features(features_offset);
  }
  result_builder.add_image_monotonic_timestamp_ns(
      image.monotonic_timestamp_ns());
  result_builder.add_camera_calibration(camera_calibration_offset);

  // TODO<Jim>: Need to add target point computed from matches and
  // mapped by homography
  builder.Send(result_builder.Finish());
}

void CameraReader::ProcessImage(const CameraImage &image) {
  // First, we need to extract the brightness information. This can't really be
  // fused into the beginning of the SIFT algorithm because the algorithm needs
  // to look at the base image directly. It also only takes 2ms on our images.
  // This is converting from YUYV to a grayscale image.
  cv::Mat image_mat(image.rows(), image.cols(), CV_8U);
  CHECK(image_mat.isContinuous());
  const int number_pixels = image.rows() * image.cols();
  for (int i = 0; i < number_pixels; ++i) {
    reinterpret_cast<uint8_t *>(image_mat.data)[i] =
        image.data()->data()[i * 2];
  }

  // Next, grab the features from the image.
  std::vector<cv::KeyPoint> keypoints;

  cv::Mat descriptors;
  if (!FLAGS_skip_sift) {
    sift_->detectAndCompute(image_mat, cv::noArray(), keypoints, descriptors);
  }

  struct PerImageMatches {
    std::vector<std::vector<cv::DMatch>> matches;
    std::vector<cv::Point3f> training_points_3d;
    std::vector<cv::Point2f> query_points;
    std::vector<cv::Point2f> training_points;
    cv::Mat homography;
  };
  std::vector<PerImageMatches> per_image_matches(number_training_images());

  for (int image_idx = 0; image_idx < number_training_images(); ++image_idx) {
    // Then, match those features against our training data.
    std::vector<std::vector<cv::DMatch>> matches;
    if (!FLAGS_skip_sift) {
      matchers_[image_idx].knnMatch(/* queryDescriptors */ descriptors, matches,
                                    /* k */ 2);
    }

    // Pull out the good matches which we want for each image.
    // Discard the bad matches per Lowe's ratio test.
    // (Lowe originally proposed 0.7 ratio, but 0.75 was later proposed as a
    // better option.  We'll go with the more conservative (fewer, better
    // matches) for now).
    for (const std::vector<cv::DMatch> &match : matches) {
      CHECK_EQ(2u, match.size());
      CHECK_LE(match[0].distance, match[1].distance);
      CHECK_EQ(match[0].imgIdx, 0);
      CHECK_EQ(match[1].imgIdx, 0);
      CHECK_EQ(match[0].queryIdx, match[1].queryIdx);
      if (!(match[0].distance < 0.7 * match[1].distance)) {
        continue;
      }

      const int training_image = image_idx;
      CHECK_LT(training_image, static_cast<int>(per_image_matches.size()));
      PerImageMatches *const per_image = &per_image_matches[training_image];
      per_image->matches.push_back(match);
      per_image->matches.back()[0].imgIdx = image_idx;
      per_image->matches.back()[1].imgIdx = image_idx;
      per_image->training_points.push_back(
          Training2dPoint(training_image, match[0].trainIdx));
      per_image->training_points_3d.push_back(
          Training3dPoint(training_image, match[0].trainIdx));

      const cv::KeyPoint &keypoint = keypoints[match[0].queryIdx];
      per_image->query_points.push_back(keypoint.pt);
    }
  }

  // The minimum number of matches in a training image for us to use it.
  static constexpr int kMinimumMatchCount = 10;
  std::vector<cv::Mat> camera_target_list;
  std::vector<cv::Mat> field_camera_list;

  // Rebuild the matches and store them here
  std::vector<std::vector<cv::DMatch>> all_good_matches;
  // Build list of target point and radius for each good match
  std::vector<cv::Point2f> target_point_vector;
  std::vector<float> target_radius_vector;
  std::vector<int> training_image_indices;

  // Iterate through matches for each training image
  for (size_t i = 0; i < per_image_matches.size(); ++i) {
    const PerImageMatches &per_image = per_image_matches[i];

    VLOG(2) << "Number of matches to start: " << per_image.matches.size()
            << "\n";
    // If we don't have enough matches to start, skip this set of matches
    if (per_image.matches.size() < kMinimumMatchCount) {
      continue;
    }

    // Use homography to determine which matches make sense physically
    cv::Mat mask;
    cv::Mat homography =
        cv::findHomography(per_image.training_points, per_image.query_points,
                           CV_RANSAC, 3.0, mask);

    // If mask doesn't have enough leftover matches, skip these matches
    if (cv::countNonZero(mask) < kMinimumMatchCount) {
      continue;
    }

    VLOG(2) << "Number of matches after homography: " << cv::countNonZero(mask)
            << "\n";

    // Fill our match info for each good match based on homography result
    PerImageMatches per_image_good_match;
    CHECK_EQ(per_image.training_points.size(),
             (unsigned long)mask.size().height);
    for (size_t j = 0; j < per_image.matches.size(); j++) {
      // Skip if we masked out by homography
      if (mask.at<uchar>(0, j) != 1) {
        continue;
      }

      // Add this to our collection of all matches that passed our criteria
      all_good_matches.push_back(per_image.matches[j]);

      // Fill out the data for matches per image that made it past
      // homography check, for later use
      per_image_good_match.matches.push_back(per_image.matches[j]);
      per_image_good_match.training_points.push_back(
          per_image.training_points[j]);
      per_image_good_match.training_points_3d.push_back(
          per_image.training_points_3d[j]);
      per_image_good_match.query_points.push_back(per_image.query_points[j]);
    }

    // Returns from opencv are doubles (CV_64F), which don't play well
    // with our floats
    homography.convertTo(homography, CV_32F);
    per_image_good_match.homography = homography;

    CHECK_GT(per_image_good_match.matches.size(), 0u);

    // Collect training target location, so we can map it to matched image
    cv::Point2f target_point;
    float target_radius;
    TargetLocation(i, target_point, target_radius);

    // Store target_point in vector for use by perspectiveTransform
    std::vector<cv::Point2f> src_target_pt;
    src_target_pt.push_back(target_point);
    std::vector<cv::Point2f> query_target_pt;

    cv::perspectiveTransform(src_target_pt, query_target_pt, homography);

    float query_target_radius =
        target_radius *
        abs(homography.at<float>(0, 0) + homography.at<float>(1, 1)) / 2.;

    CHECK_EQ(query_target_pt.size(), 1u);
    target_point_vector.push_back(query_target_pt[0]);
    target_radius_vector.push_back(query_target_radius);

    // Pose transformations (rotations and translations) for various
    // coordinate frames.  R_X_Y_vec is the Rodrigues (angle-axis)
    // representation of the 3x3 rotation R_X_Y from frame X to frame Y

    // Tranform from camera to target frame
    cv::Mat R_camera_target_vec, R_camera_target, T_camera_target;
    // Tranform from camera to field origin (global) reference frame
    cv::Mat R_camera_field_vec, R_camera_field, T_camera_field;
    // Inverse of camera to field-- defines location of camera in
    // global (field) reference frame
    cv::Mat R_field_camera_vec, R_field_camera, T_field_camera;

    // Using the previous pose helps to stabilize the estimate, since
    // it's sometimes bouncing between two possible poses.  Putting it
    // near the previous pose helps it converge to the previous pose
    // estimate (assuming it's valid).
    if (FLAGS_use_prev_pose) {
      R_camera_field_vec = prev_R_camera_field_vec_;
      T_camera_field = prev_T_camera_field_;
    }

    // Compute the pose of the camera (global origin relative to camera)
    if (FLAGS_ransac_pose) {
      // RANSAC computation is designed to be more robust to outliers.
      // But, we found it bounces around a lot, even with identical points
      cv::solvePnPRansac(per_image_good_match.training_points_3d,
                         per_image_good_match.query_points, CameraIntrinsics(),
                         CameraDistCoeffs(), R_camera_field_vec, T_camera_field,
                         FLAGS_use_prev_pose);
    } else {
      // ITERATIVE mode is potentially less robust to outliers, but we
      // found it to be more stable
      //
      cv::solvePnP(per_image_good_match.training_points_3d,
                   per_image_good_match.query_points, CameraIntrinsics(),
                   CameraDistCoeffs(), R_camera_field_vec, T_camera_field,
                   FLAGS_use_prev_pose, CV_ITERATIVE);
    }

    prev_R_camera_field_vec_ = R_camera_field_vec;
    prev_T_camera_field_ = T_camera_field;

    CHECK_EQ(cv::Size(1, 3), T_camera_field.size());

    // Convert to float32's (from float64) to be compatible with the rest
    R_camera_field_vec.convertTo(R_camera_field_vec, CV_32F);
    T_camera_field.convertTo(T_camera_field, CV_32F);

    // Get matrix version of R_camera_field
    cv::Rodrigues(R_camera_field_vec, R_camera_field);
    CHECK_EQ(cv::Size(3, 3), R_camera_field.size());

    // Compute H_field_camera = H_camera_field^-1
    R_field_camera = R_camera_field.t();
    T_field_camera = -R_field_camera * (T_camera_field);

    // Extract the field_target transformation
    const cv::Mat H_field_target(4, 4, CV_32F,
                                 const_cast<void *>(static_cast<const void *>(
                                     FieldToTarget(i)->data()->data())));

    training_image_indices.push_back(i);

    const cv::Mat R_field_target =
        H_field_target(cv::Range(0, 3), cv::Range(0, 3));
    const cv::Mat T_field_target =
        H_field_target(cv::Range(0, 3), cv::Range(3, 4));

    // Use it to get the relative pose from camera to target
    R_camera_target = R_camera_field * (R_field_target);
    T_camera_target = R_camera_field * (T_field_target) + T_camera_field;

    // Set H_camera_target
    {
      CHECK_EQ(cv::Size(3, 3), R_camera_target.size());
      CHECK_EQ(cv::Size(1, 3), T_camera_target.size());
      cv::Mat H_camera_target = cv::Mat::zeros(4, 4, CV_32F);
      R_camera_target.copyTo(H_camera_target(cv::Range(0, 3), cv::Range(0, 3)));
      T_camera_target.copyTo(H_camera_target(cv::Range(0, 3), cv::Range(3, 4)));
      H_camera_target.at<float>(3, 3) = 1;
      CHECK(H_camera_target.isContinuous());
      camera_target_list.push_back(H_camera_target.clone());
    }

    // Set H_field_camera
    {
      CHECK_EQ(cv::Size(3, 3), R_field_camera.size());
      CHECK_EQ(cv::Size(1, 3), T_field_camera.size());
      cv::Mat H_field_camera = cv::Mat::zeros(4, 4, CV_32F);
      R_field_camera.copyTo(H_field_camera(cv::Range(0, 3), cv::Range(0, 3)));
      T_field_camera.copyTo(H_field_camera(cv::Range(0, 3), cv::Range(3, 4)));
      H_field_camera.at<float>(3, 3) = 1;
      CHECK(H_field_camera.isContinuous());
      field_camera_list.push_back(H_field_camera.clone());
    }
  }
  // Now, send our two messages-- one large, with details for remote
  // debugging(features), and one smaller
  SendImageMatchResult(image, keypoints, descriptors, all_good_matches,
                       camera_target_list, field_camera_list,
                       target_point_vector, target_radius_vector,
                       training_image_indices, &detailed_result_sender_, true);
  SendImageMatchResult(image, keypoints, descriptors, all_good_matches,
                       camera_target_list, field_camera_list,
                       target_point_vector, target_radius_vector,
                       training_image_indices, &result_sender_, false);
}

void CameraReader::ReadImage() {
  if (!reader_->ReadLatestImage()) {
    if (!FLAGS_skip_sift) {
      LOG(INFO) << "No image, sleeping";
    }
    read_image_timer_->Setup(event_loop_->monotonic_now() +
                             std::chrono::milliseconds(10));
    return;
  }

  ProcessImage(reader_->LatestImage());

  reader_->SendLatestImage();
  read_image_timer_->Setup(event_loop_->monotonic_now());
}

flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<sift::ImageMatch>>>
CameraReader::PackImageMatches(
    flatbuffers::FlatBufferBuilder *fbb,
    const std::vector<std::vector<cv::DMatch>> &matches) {
  // First, we need to pull out all the matches for each image. Might as well
  // build up the Match tables at the same time.
  std::vector<std::vector<sift::Match>> per_image_matches(
      number_training_images());
  for (const std::vector<cv::DMatch> &image_matches : matches) {
    CHECK_GT(image_matches.size(), 0u);
    // We're only using the first of the two matches
    const cv::DMatch &image_match = image_matches[0];
    CHECK_LT(image_match.imgIdx, number_training_images());
    per_image_matches[image_match.imgIdx].emplace_back();
    sift::Match *const match = &per_image_matches[image_match.imgIdx].back();
    match->mutate_query_feature(image_match.queryIdx);
    match->mutate_train_feature(image_match.trainIdx);
    match->mutate_distance(image_match.distance);
  }

  // Then, we need to build up each ImageMatch table.
  std::vector<flatbuffers::Offset<sift::ImageMatch>> image_match_tables;
  for (size_t i = 0; i < per_image_matches.size(); ++i) {
    const std::vector<sift::Match> &this_image_matches = per_image_matches[i];
    if (this_image_matches.empty()) {
      continue;
    }
    const auto vector_offset = fbb->CreateVectorOfStructs(this_image_matches);
    sift::ImageMatch::Builder image_builder(*fbb);
    image_builder.add_train_image(i);
    image_builder.add_matches(vector_offset);
    image_match_tables.emplace_back(image_builder.Finish());
  }

  return fbb->CreateVector(image_match_tables);
}

flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<sift::Feature>>>
CameraReader::PackFeatures(flatbuffers::FlatBufferBuilder *fbb,
                           const std::vector<cv::KeyPoint> &keypoints,
                           const cv::Mat &descriptors) {
  const int number_features = keypoints.size();
  CHECK_EQ(descriptors.rows, number_features);
  if (number_features != 0) {
    CHECK_EQ(descriptors.cols, 128);
  }
  std::vector<flatbuffers::Offset<sift::Feature>> features_vector(
      number_features);
  for (int i = 0; i < number_features; ++i) {
    const auto submat =
        descriptors(cv::Range(i, i + 1), cv::Range(0, descriptors.cols));
    CHECK(submat.isContinuous());
    flatbuffers::Offset<flatbuffers::Vector<uint8_t>> descriptor_offset;
    {
      uint8_t *data;
      descriptor_offset = fbb->CreateUninitializedVector(128, &data);
      submat.convertTo(
          cv::Mat(1, descriptors.cols, CV_8U, static_cast<void *>(data)),
          CV_8U);
    }
    sift::Feature::Builder feature_builder(*fbb);
    feature_builder.add_descriptor(descriptor_offset);
    feature_builder.add_x(keypoints[i].pt.x);
    feature_builder.add_y(keypoints[i].pt.y);
    feature_builder.add_size(keypoints[i].size);
    feature_builder.add_angle(keypoints[i].angle);
    feature_builder.add_response(keypoints[i].response);
    feature_builder.add_octave(keypoints[i].octave);
    CHECK_EQ(-1, keypoints[i].class_id)
        << ": Not sure what to do with a class id";
    features_vector[i] = feature_builder.Finish();
  }
  return fbb->CreateVector(features_vector);
}

void CameraReaderMain() {
  aos::FlatbufferDetachedBuffer<aos::Configuration> config =
      aos::configuration::ReadConfig(FLAGS_config);

  const aos::FlatbufferSpan<sift::TrainingData> training_data(
      SiftTrainingData());
  CHECK(training_data.Verify());

  const auto index_params = cv::makePtr<cv::flann::IndexParams>();
  index_params->setAlgorithm(cvflann::FLANN_INDEX_KDTREE);
  index_params->setInt("trees", 5);
  const auto search_params =
      cv::makePtr<cv::flann::SearchParams>(/* checks */ 50);
  cv::FlannBasedMatcher matcher(index_params, search_params);

  aos::ShmEventLoop event_loop(&config.message());

  // First, log the data for future reference.
  {
    aos::Sender<sift::TrainingData> training_data_sender =
        event_loop.MakeSender<sift::TrainingData>("/camera");
    training_data_sender.Send(training_data);
  }

  V4L2Reader v4l2_reader(&event_loop, "/dev/video0");
  CameraReader camera_reader(&event_loop, &training_data.message(),
                             &v4l2_reader, index_params, search_params);

  event_loop.Run();
}

}  // namespace
}  // namespace vision
}  // namespace frc971

int main(int argc, char **argv) {
  aos::InitGoogle(&argc, &argv);
  frc971::vision::CameraReaderMain();
}