Function parseImagesBinary

src/parsers/images.ts:31–92

(
  buffer: ArrayBuffer,
  skipPoints2D: boolean = false
)

 *                       For large datasets (1GB+ images.bin), this reduces memory by ~80-90%.
 */
export function parseImagesBinary(
  buffer: ArrayBuffer,
  skipPoints2D: boolean = false
): Map<number, Image> {
  const reader = new BinaryReader(buffer);
  const images = new Map<number, Image>();

  const numImages = reader.readUint64AsNumber();

  for (let i = 0; i < numImages; i++) {
    const imageId = reader.readUint32();

    // Read quaternion (qw, qx, qy, qz)
    const qvec: [number, number, number, number] = [
      reader.readFloat64(),
      reader.readFloat64(),
      reader.readFloat64(),
      reader.readFloat64(),
    ];

    // Read translation (tx, ty, tz)
    const tvec: [number, number, number] = [
      reader.readFloat64(),
      reader.readFloat64(),
      reader.readFloat64(),
    ];

    const cameraId = reader.readUint32();
    const name = reader.readString();

    const numPoints2D = reader.readUint64AsNumber();
    const points2D: Point2D[] = [];

    if (skipPoints2D) {
      // Skip all points2D data: each point is 8 + 8 + 8 = 24 bytes (x, y, point3D_id)
      reader.skip(numPoints2D * 24);
    } else {
      for (let j = 0; j < numPoints2D; j++) {
        const x = reader.readFloat64();
        const y = reader.readFloat64();
        const point3DId = reader.readInt64(); // signed, -1 means not triangulated

        points2D.push({
          xy: [x, y],
          point3DId: point3DId,
        });
      }
    }

    images.set(imageId, {
      imageId,
      qvec,
      tvec,
      cameraId,
      name,
      points2D,
      numPoints2D, // Always store the count
    });