Hi there, your values ​​are quite high, so try lowering it and try again on more samples.

In our final implementation we don't even use bounding boxes and therefore NMS anymore.

V1.0

22841 frames, 16 seconds, 1411.40frame/s, success rate=75.2%

V1.1

22841 frames, 15seconds, 1485.71frame/s, success rate=46.8%

def detect_nsfw(vision_frame : VisionFrame) -> List[Score]:
    nsfw_scores = []
    model_size = get_model_options().get('size')
    temp_vision_frame = resize_frame_resolution(vision_frame, model_size)
    detect_vision_frame = prepare_detect_frame(temp_vision_frame)
    detection = forward(detect_vision_frame)
    detection = numpy.squeeze(detection).T
    nsfw_scores_raw = numpy.amax(detection[:, 4:], axis = 1)
    keep_indices = numpy.where(nsfw_scores_raw > 0.2)[0]

    if numpy.any(keep_indices):
        nsfw_scores_raw = nsfw_scores_raw[keep_indices]
        nsfw_scores = nsfw_scores_raw.ravel().tolist()

    return nsfw_scores


def forward(vision_frame : VisionFrame) -> Detection:
    content_analyser = get_inference_pool().get('content_analyser')

    with conditional_thread_semaphore():
        detection = content_analyser.run(None,
        {
            'input': vision_frame
        })

    return detection


def prepare_detect_frame(temp_vision_frame : VisionFrame) -> VisionFrame:
    model_size = get_model_options().get('size')
    detect_vision_frame = numpy.zeros((model_size[0], model_size[1], 3))
    detect_vision_frame[:temp_vision_frame.shape[0], :temp_vision_frame.shape[1], :] = temp_vision_frame
    detect_vision_frame = detect_vision_frame / 255.0
    detect_vision_frame = numpy.expand_dims(detect_vision_frame.transpose(2, 0, 1), axis = 0).astype(numpy.float32)
    return detect_vision_frame


def resize_frame_resolution(vision_frame : VisionFrame, max_resolution : Resolution) -> VisionFrame:
    height, width = vision_frame.shape[:2]
    max_width, max_height = max_resolution

    if height > max_height or width > max_width:
        scale = min(max_height / height, max_width / width)
        new_width = int(width * scale)
        new_height = int(height * scale)
        return cv2.resize(vision_frame, (new_width, new_height))
    return vision_frame